From df4abd8b14ff823f060f66acde34713da013d062 Mon Sep 17 00:00:00 2001 From: Conrad Ludgate Date: Tue, 14 Jan 2025 12:53:32 +0000 Subject: [PATCH 01/31] fix: force-refresh azure identity token (#10378) ## Problem Because of https://github.com/Azure/azure-sdk-for-rust/issues/1739, our identity token file was not being refreshed. This caused our uploads to start failing when the storage token expired. ## Summary of changes Drop and recreate the remote storage config every time we upload in order to force reload the identity token file. --- proxy/src/context/parquet.rs | 73 +++++++++++++++++++++++------------- 1 file changed, 46 insertions(+), 27 deletions(-) diff --git a/proxy/src/context/parquet.rs b/proxy/src/context/parquet.rs index 5f65b17374..d7ffff0483 100644 --- a/proxy/src/context/parquet.rs +++ b/proxy/src/context/parquet.rs @@ -187,10 +187,6 @@ pub async fn worker( let rx = futures::stream::poll_fn(move |cx| rx.poll_recv(cx)); let rx = rx.map(RequestData::from); - let storage = GenericRemoteStorage::from_config(&remote_storage_config) - .await - .context("remote storage init")?; - let properties = WriterProperties::builder() .set_data_page_size_limit(config.parquet_upload_page_size) .set_compression(config.parquet_upload_compression); @@ -224,18 +220,18 @@ pub async fn worker( let rx_disconnect = futures::stream::poll_fn(move |cx| rx_disconnect.poll_recv(cx)); let rx_disconnect = rx_disconnect.map(RequestData::from); - let storage_disconnect = - GenericRemoteStorage::from_config(&disconnect_events_storage_config) - .await - .context("remote storage for disconnect events init")?; let parquet_config_disconnect = parquet_config.clone(); tokio::try_join!( - worker_inner(storage, rx, parquet_config), - worker_inner(storage_disconnect, rx_disconnect, parquet_config_disconnect) + worker_inner(remote_storage_config, rx, parquet_config), + worker_inner( + disconnect_events_storage_config, + rx_disconnect, + parquet_config_disconnect + ) ) .map(|_| ()) } else { - worker_inner(storage, rx, parquet_config).await + worker_inner(remote_storage_config, rx, parquet_config).await } } @@ -251,18 +247,32 @@ struct ParquetConfig { test_remote_failures: u64, } +impl ParquetConfig { + async fn storage( + &self, + storage_config: &RemoteStorageConfig, + ) -> anyhow::Result { + let storage = GenericRemoteStorage::from_config(storage_config) + .await + .context("remote storage init")?; + + #[cfg(any(test, feature = "testing"))] + if self.test_remote_failures > 0 { + return Ok(GenericRemoteStorage::unreliable_wrapper( + storage, + self.test_remote_failures, + )); + } + + Ok(storage) + } +} + async fn worker_inner( - storage: GenericRemoteStorage, + storage_config: RemoteStorageConfig, rx: impl Stream, config: ParquetConfig, ) -> anyhow::Result<()> { - #[cfg(any(test, feature = "testing"))] - let storage = if config.test_remote_failures > 0 { - GenericRemoteStorage::unreliable_wrapper(storage, config.test_remote_failures) - } else { - storage - }; - let mut rx = std::pin::pin!(rx); let mut rows = Vec::with_capacity(config.rows_per_group); @@ -285,7 +295,7 @@ async fn worker_inner( } if len > config.file_size || force { last_upload = time::Instant::now(); - let file = upload_parquet(w, len, &storage).await?; + let file = upload_parquet(w, len, &storage_config, &config).await?; w = SerializedFileWriter::new(file, schema.clone(), config.propeties.clone())?; len = 0; } @@ -298,7 +308,7 @@ async fn worker_inner( } if !w.flushed_row_groups().is_empty() { - let _rtchk: Writer = upload_parquet(w, len, &storage).await?; + let _rtchk: Writer = upload_parquet(w, len, &storage_config, &config).await?; } Ok(()) @@ -340,7 +350,8 @@ where async fn upload_parquet( mut w: SerializedFileWriter>, len: i64, - storage: &GenericRemoteStorage, + storage_config: &RemoteStorageConfig, + config: &ParquetConfig, ) -> anyhow::Result> { let len_uncompressed = w .flushed_row_groups() @@ -377,6 +388,15 @@ async fn upload_parquet( size, compression, "uploading request parquet file" ); + // A bug in azure-sdk means that the identity-token-file that expires after + // 1 hour is not refreshed. This identity-token is used to fetch the actual azure storage + // tokens that last for 24 hours. After this 24 hour period, azure-sdk tries to refresh + // the storage token, but the identity token has now expired. + // + // + // To work around this, we recreate the storage every time. + let storage = config.storage(storage_config).await?; + let year = now.year(); let month = now.month(); let day = now.day(); @@ -431,8 +451,8 @@ mod tests { use rand::rngs::StdRng; use rand::{Rng, SeedableRng}; use remote_storage::{ - GenericRemoteStorage, RemoteStorageConfig, RemoteStorageKind, S3Config, - DEFAULT_MAX_KEYS_PER_LIST_RESPONSE, DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT, + RemoteStorageConfig, RemoteStorageKind, S3Config, DEFAULT_MAX_KEYS_PER_LIST_RESPONSE, + DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT, }; use tokio::sync::mpsc; use tokio::time; @@ -559,12 +579,11 @@ mod tests { timeout: std::time::Duration::from_secs(120), small_timeout: std::time::Duration::from_secs(30), }; - let storage = GenericRemoteStorage::from_config(&remote_storage_config) + + worker_inner(remote_storage_config, rx, config) .await .unwrap(); - worker_inner(storage, rx, config).await.unwrap(); - let mut files = WalkDir::new(tmpdir.as_std_path()) .into_iter() .filter_map(|entry| entry.ok()) From 9bdb14c1c0b3ef1082ce68f1c54d4547393da362 Mon Sep 17 00:00:00 2001 From: "Alex Chi Z." <4198311+skyzh@users.noreply.github.com> Date: Tue, 14 Jan 2025 10:27:48 -0500 Subject: [PATCH 02/31] fix(pageserver): ensure initial image layers have correct key ranges (#10374) ## Problem Discovered during the relation dir refactor work. If we do not create images as in this patch, we would get two set of image layers: ``` 0000...METADATA_KEYS 0000...REL_KEYS ``` They overlap at the same LSN and would cause data loss for relation keys. This doesn't happen in prod because initial image layer generation is never called, but better to be fixed to avoid future issues with the reldir refactors. ## Summary of changes * Consolidate create_image_layers call into a single one. --------- Signed-off-by: Alex Chi Z --- pageserver/src/tenant/timeline.rs | 43 +++++++++++++++---------------- 1 file changed, 21 insertions(+), 22 deletions(-) diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs index f7227efeba..741b214a73 100644 --- a/pageserver/src/tenant/timeline.rs +++ b/pageserver/src/tenant/timeline.rs @@ -3781,36 +3781,35 @@ impl Timeline { return Err(FlushLayerError::Cancelled); } - let mut layers_to_upload = Vec::new(); - layers_to_upload.extend( - self.create_image_layers( - &rel_partition, - self.initdb_lsn, - ImageLayerCreationMode::Initial, - ctx, - ) - .await?, - ); + // Ensure that we have a single call to `create_image_layers` with a combined dense keyspace. + // So that the key ranges don't overlap. + let mut partitions = KeyPartitioning::default(); + partitions.parts.extend(rel_partition.parts); if !metadata_partition.parts.is_empty() { assert_eq!( metadata_partition.parts.len(), 1, "currently sparse keyspace should only contain a single metadata keyspace" ); - layers_to_upload.extend( - self.create_image_layers( - // Safety: create_image_layers treat sparse keyspaces differently that it does not scan - // every single key within the keyspace, and therefore, it's safe to force converting it - // into a dense keyspace before calling this function. - &metadata_partition.into_dense(), - self.initdb_lsn, - ImageLayerCreationMode::Initial, - ctx, - ) - .await?, - ); + // Safety: create_image_layers treat sparse keyspaces differently that it does not scan + // every single key within the keyspace, and therefore, it's safe to force converting it + // into a dense keyspace before calling this function. + partitions + .parts + .extend(metadata_partition.into_dense().parts); } + let mut layers_to_upload = Vec::new(); + layers_to_upload.extend( + self.create_image_layers( + &partitions, + self.initdb_lsn, + ImageLayerCreationMode::Initial, + ctx, + ) + .await?, + ); + (layers_to_upload, None) } else { // Normal case, write out a L0 delta layer file. From 2466a2f97763aff5b84253a242909fc19bc995a2 Mon Sep 17 00:00:00 2001 From: Christian Schwarz Date: Tue, 14 Jan 2025 16:28:01 +0100 Subject: [PATCH 03/31] page_service: throttle individual requests instead of the batched request (#10353) ## Problem Before this PR, the pagestream throttle was applied weighted on a per-batch basis. This had several problems: 1. The throttle occurence counters were only bumped by `1` instead of `batch_size`. 2. The throttle wait time aggregator metric only counted one wait time, irrespective of `batch_size`. That makes sense in some ways of looking at it but not in others. 3. If the last request in the batch runs into the throttle, the other requests in the batch are also throttled, i.e., over-throttling happens (theoretical, didn't measure it in practice). ## Solution It occured to me that we can simply push the throttling upwards into `pagestream_read_message`. This has the added benefit that in pipeline mode, the `executor` stage will, if it is idle, steal whatever requests already made it into the `spsc_fold` and execute them; before this change, that was not the case - the throttling happened in the `executor` stage instead of the `batcher` stage. ## Code Changes There are two changes in this PR: 1. Lifting up the throttling into the `pagestream_read_message` method. 2. Move the throttling metrics out of the `Throttle` type into `SmgrOpMetrics`. Unlike the other smgr metrics, throttling is per-tenant, hence the Arc. 3. Refactor the `SmgrOpTimer` implementation to account for the new observation states, and simplify its design. 4. Drive-by-fix flush time metrics. It was using the same `now` in the `observe_guard` every time. The `SmgrOpTimer` is now a state machine. Each observation point moves the state machine forward. If a timer object is dropped early some "pair"-like metrics still require an increment or observation. That's done in the Drop implementation, by driving the state machine to completion. --- pageserver/src/metrics.rs | 367 +++++++++++------------ pageserver/src/page_service.rs | 139 +++++---- pageserver/src/tenant.rs | 14 +- pageserver/src/tenant/throttle.rs | 45 +-- pageserver/src/tenant/timeline.rs | 8 +- pageserver/src/tenant/timeline/delete.rs | 1 + 6 files changed, 284 insertions(+), 290 deletions(-) diff --git a/pageserver/src/metrics.rs b/pageserver/src/metrics.rs index 5b8419fda9..5b1cbbad63 100644 --- a/pageserver/src/metrics.rs +++ b/pageserver/src/metrics.rs @@ -1224,117 +1224,189 @@ pub(crate) struct SmgrOpTimerInner { global_flush_in_progress_micros: IntCounter, per_timeline_flush_in_progress_micros: IntCounter, + throttling: Arc, + timings: SmgrOpTimerState, } +/// The stages of request processing are represented by the enum variants. +/// Used as part of [`SmgrOpTimerInner::timings`]. +/// +/// Request processing calls into the `SmgrOpTimer::observe_*` methods at the +/// transition points. +/// These methods bump relevant counters and then update [`SmgrOpTimerInner::timings`] +/// to the next state. +/// +/// Each request goes through every stage, in all configurations. +/// #[derive(Debug)] enum SmgrOpTimerState { Received { + // In the future, we may want to track the full time the request spent + // inside pageserver process (time spent in kernel buffers can't be tracked). + // `received_at` would be used for that. + #[allow(dead_code)] received_at: Instant, }, - ThrottleDoneExecutionStarting { - received_at: Instant, + Throttling { throttle_started_at: Instant, - started_execution_at: Instant, }, + Batching { + throttle_done_at: Instant, + }, + Executing { + execution_started_at: Instant, + }, + Flushing, + // NB: when adding observation points, remember to update the Drop impl. } +// NB: when adding observation points, remember to update the Drop impl. +impl SmgrOpTimer { + /// See [`SmgrOpTimerState`] for more context. + pub(crate) fn observe_throttle_start(&mut self, at: Instant) { + let Some(inner) = self.0.as_mut() else { + return; + }; + let SmgrOpTimerState::Received { received_at: _ } = &mut inner.timings else { + return; + }; + inner.throttling.count_accounted_start.inc(); + inner.timings = SmgrOpTimerState::Throttling { + throttle_started_at: at, + }; + } + + /// See [`SmgrOpTimerState`] for more context. + pub(crate) fn observe_throttle_done(&mut self, throttle: ThrottleResult) { + let Some(inner) = self.0.as_mut() else { + return; + }; + let SmgrOpTimerState::Throttling { + throttle_started_at, + } = &inner.timings + else { + return; + }; + inner.throttling.count_accounted_finish.inc(); + match throttle { + ThrottleResult::NotThrottled { end } => { + inner.timings = SmgrOpTimerState::Batching { + throttle_done_at: end, + }; + } + ThrottleResult::Throttled { end } => { + // update metrics + inner.throttling.count_throttled.inc(); + inner + .throttling + .wait_time + .inc_by((end - *throttle_started_at).as_micros().try_into().unwrap()); + // state transition + inner.timings = SmgrOpTimerState::Batching { + throttle_done_at: end, + }; + } + } + } + + /// See [`SmgrOpTimerState`] for more context. + pub(crate) fn observe_execution_start(&mut self, at: Instant) { + let Some(inner) = self.0.as_mut() else { + return; + }; + let SmgrOpTimerState::Batching { throttle_done_at } = &inner.timings else { + return; + }; + // update metrics + let batch = at - *throttle_done_at; + inner.global_batch_wait_time.observe(batch.as_secs_f64()); + inner + .per_timeline_batch_wait_time + .observe(batch.as_secs_f64()); + // state transition + inner.timings = SmgrOpTimerState::Executing { + execution_started_at: at, + } + } + + /// For all but the first caller, this is a no-op. + /// The first callers receives Some, subsequent ones None. + /// + /// See [`SmgrOpTimerState`] for more context. + pub(crate) fn observe_execution_end_flush_start( + &mut self, + at: Instant, + ) -> Option { + // NB: unlike the other observe_* methods, this one take()s. + #[allow(clippy::question_mark)] // maintain similar code pattern. + let Some(mut inner) = self.0.take() else { + return None; + }; + let SmgrOpTimerState::Executing { + execution_started_at, + } = &inner.timings + else { + return None; + }; + // update metrics + let execution = at - *execution_started_at; + inner + .global_execution_latency_histo + .observe(execution.as_secs_f64()); + if let Some(per_timeline_execution_latency_histo) = + &inner.per_timeline_execution_latency_histo + { + per_timeline_execution_latency_histo.observe(execution.as_secs_f64()); + } + + // state transition + inner.timings = SmgrOpTimerState::Flushing; + + // return the flush in progress object which + // will do the remaining metrics updates + let SmgrOpTimerInner { + global_flush_in_progress_micros, + per_timeline_flush_in_progress_micros, + .. + } = inner; + Some(SmgrOpFlushInProgress { + flush_started_at: at, + global_micros: global_flush_in_progress_micros, + per_timeline_micros: per_timeline_flush_in_progress_micros, + }) + } +} + +/// The last stage of request processing is serializing and flushing the request +/// into the TCP connection. We want to make slow flushes observable +/// _while they are occuring_, so this struct provides a wrapper method [`Self::measure`] +/// to periodically bump the metric. +/// +/// If in the future we decide that we're not interested in live updates, we can +/// add another `observe_*` method to [`SmgrOpTimer`], follow the existing pattern there, +/// and remove this struct from the code base. pub(crate) struct SmgrOpFlushInProgress { flush_started_at: Instant, global_micros: IntCounter, per_timeline_micros: IntCounter, } -impl SmgrOpTimer { - pub(crate) fn observe_throttle_done_execution_starting(&mut self, throttle: &ThrottleResult) { - let inner = self.0.as_mut().expect("other public methods consume self"); - match (&mut inner.timings, throttle) { - (SmgrOpTimerState::Received { received_at }, throttle) => match throttle { - ThrottleResult::NotThrottled { start } => { - inner.timings = SmgrOpTimerState::ThrottleDoneExecutionStarting { - received_at: *received_at, - throttle_started_at: *start, - started_execution_at: *start, - }; - } - ThrottleResult::Throttled { start, end } => { - inner.timings = SmgrOpTimerState::ThrottleDoneExecutionStarting { - received_at: *start, - throttle_started_at: *start, - started_execution_at: *end, - }; - } - }, - (x, _) => panic!("called in unexpected state: {x:?}"), - } - } - - pub(crate) fn observe_smgr_op_completion_and_start_flushing(mut self) -> SmgrOpFlushInProgress { - let (flush_start, inner) = self - .smgr_op_end() - .expect("this method consume self, and the only other caller is drop handler"); - let SmgrOpTimerInner { - global_flush_in_progress_micros, - per_timeline_flush_in_progress_micros, - .. - } = inner; - SmgrOpFlushInProgress { - flush_started_at: flush_start, - global_micros: global_flush_in_progress_micros, - per_timeline_micros: per_timeline_flush_in_progress_micros, - } - } - - /// Returns `None`` if this method has already been called, `Some` otherwise. - fn smgr_op_end(&mut self) -> Option<(Instant, SmgrOpTimerInner)> { - let inner = self.0.take()?; - - let now = Instant::now(); - - let batch; - let execution; - let throttle; - match inner.timings { - SmgrOpTimerState::Received { received_at } => { - batch = (now - received_at).as_secs_f64(); - // TODO: use label for dropped requests. - // This is quite rare in practice, only during tenant/pageservers shutdown. - throttle = Duration::ZERO; - execution = Duration::ZERO.as_secs_f64(); - } - SmgrOpTimerState::ThrottleDoneExecutionStarting { - received_at, - throttle_started_at, - started_execution_at, - } => { - batch = (throttle_started_at - received_at).as_secs_f64(); - throttle = started_execution_at - throttle_started_at; - execution = (now - started_execution_at).as_secs_f64(); - } - } - - // update time spent in batching - inner.global_batch_wait_time.observe(batch); - inner.per_timeline_batch_wait_time.observe(batch); - - // time spent in throttle metric is updated by throttle impl - let _ = throttle; - - // update metrics for execution latency - inner.global_execution_latency_histo.observe(execution); - if let Some(per_timeline_execution_latency_histo) = - &inner.per_timeline_execution_latency_histo - { - per_timeline_execution_latency_histo.observe(execution); - } - - Some((now, inner)) - } -} - impl Drop for SmgrOpTimer { fn drop(&mut self) { - self.smgr_op_end(); + // In case of early drop, update any of the remaining metrics with + // observations so that (started,finished) counter pairs balance out + // and all counters on the latency path have the the same number of + // observations. + // It's technically lying and it would be better if each metric had + // a separate label or similar for cancelled requests. + // But we don't have that right now and counter pairs balancing + // out is useful when using the metrics in panels and whatnot. + let now = Instant::now(); + self.observe_throttle_start(now); + self.observe_throttle_done(ThrottleResult::NotThrottled { end: now }); + self.observe_execution_start(now); + self.observe_execution_end_flush_start(now); } } @@ -1345,12 +1417,12 @@ impl SmgrOpFlushInProgress { { let mut fut = std::pin::pin!(fut); - let now = Instant::now(); // Whenever observe_guard gets called, or dropped, // it adds the time elapsed since its last call to metrics. // Last call is tracked in `now`. let mut observe_guard = scopeguard::guard( || { + let now = Instant::now(); let elapsed = now - self.flush_started_at; self.global_micros .inc_by(u64::try_from(elapsed.as_micros()).unwrap()); @@ -1393,7 +1465,6 @@ pub enum SmgrQueryType { GetSlruSegment, } -#[derive(Debug)] pub(crate) struct SmgrQueryTimePerTimeline { global_started: [IntCounter; SmgrQueryType::COUNT], global_latency: [Histogram; SmgrQueryType::COUNT], @@ -1405,6 +1476,7 @@ pub(crate) struct SmgrQueryTimePerTimeline { per_timeline_flush_in_progress_micros: IntCounter, global_batch_wait_time: Histogram, per_timeline_batch_wait_time: Histogram, + throttling: Arc, } static SMGR_QUERY_STARTED_GLOBAL: Lazy = Lazy::new(|| { @@ -1610,7 +1682,11 @@ static PAGE_SERVICE_SMGR_BATCH_WAIT_TIME_GLOBAL: Lazy = Lazy::new(|| }); impl SmgrQueryTimePerTimeline { - pub(crate) fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self { + pub(crate) fn new( + tenant_shard_id: &TenantShardId, + timeline_id: &TimelineId, + pagestream_throttle_metrics: Arc, + ) -> Self { let tenant_id = tenant_shard_id.tenant_id.to_string(); let shard_slug = format!("{}", tenant_shard_id.shard_slug()); let timeline_id = timeline_id.to_string(); @@ -1671,6 +1747,7 @@ impl SmgrQueryTimePerTimeline { per_timeline_flush_in_progress_micros, global_batch_wait_time, per_timeline_batch_wait_time, + throttling: pagestream_throttle_metrics, } } pub(crate) fn start_smgr_op(&self, op: SmgrQueryType, received_at: Instant) -> SmgrOpTimer { @@ -1686,88 +1763,24 @@ impl SmgrQueryTimePerTimeline { SmgrOpTimer(Some(SmgrOpTimerInner { global_execution_latency_histo: self.global_latency[op as usize].clone(), per_timeline_execution_latency_histo: per_timeline_latency_histo, - timings: SmgrOpTimerState::Received { received_at }, global_flush_in_progress_micros: self.global_flush_in_progress_micros.clone(), per_timeline_flush_in_progress_micros: self .per_timeline_flush_in_progress_micros .clone(), global_batch_wait_time: self.global_batch_wait_time.clone(), per_timeline_batch_wait_time: self.per_timeline_batch_wait_time.clone(), + throttling: self.throttling.clone(), + timings: SmgrOpTimerState::Received { received_at }, })) } + /// TODO: do something about this? seems odd, we have a similar call on SmgrOpTimer pub(crate) fn observe_getpage_batch_start(&self, batch_size: usize) { self.global_batch_size.observe(batch_size as f64); self.per_timeline_batch_size.observe(batch_size as f64); } } -#[cfg(test)] -mod smgr_query_time_tests { - use std::time::Instant; - - use pageserver_api::shard::TenantShardId; - use strum::IntoEnumIterator; - use utils::id::{TenantId, TimelineId}; - - // Regression test, we used hard-coded string constants before using an enum. - #[test] - fn op_label_name() { - use super::SmgrQueryType::*; - let expect: [(super::SmgrQueryType, &'static str); 5] = [ - (GetRelExists, "get_rel_exists"), - (GetRelSize, "get_rel_size"), - (GetPageAtLsn, "get_page_at_lsn"), - (GetDbSize, "get_db_size"), - (GetSlruSegment, "get_slru_segment"), - ]; - for (op, expect) in expect { - let actual: &'static str = op.into(); - assert_eq!(actual, expect); - } - } - - #[test] - fn basic() { - let ops: Vec<_> = super::SmgrQueryType::iter().collect(); - - for op in &ops { - let tenant_id = TenantId::generate(); - let timeline_id = TimelineId::generate(); - let metrics = super::SmgrQueryTimePerTimeline::new( - &TenantShardId::unsharded(tenant_id), - &timeline_id, - ); - - let get_counts = || { - let global: u64 = ops - .iter() - .map(|op| metrics.global_latency[*op as usize].get_sample_count()) - .sum(); - ( - global, - metrics.per_timeline_getpage_latency.get_sample_count(), - ) - }; - - let (pre_global, pre_per_tenant_timeline) = get_counts(); - assert_eq!(pre_per_tenant_timeline, 0); - - let timer = metrics.start_smgr_op(*op, Instant::now()); - drop(timer); - - let (post_global, post_per_tenant_timeline) = get_counts(); - if matches!(op, super::SmgrQueryType::GetPageAtLsn) { - // getpage ops are tracked per-timeline, others aren't - assert_eq!(post_per_tenant_timeline, 1); - } else { - assert_eq!(post_per_tenant_timeline, 0); - } - assert!(post_global > pre_global); - } - } -} - // keep in sync with control plane Go code so that we can validate // compute's basebackup_ms metric with our perspective in the context of SLI/SLO. static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| { @@ -3563,9 +3576,7 @@ pub(crate) mod tenant_throttling { use once_cell::sync::Lazy; use utils::shard::TenantShardId; - use crate::tenant::{self}; - - struct GlobalAndPerTenantIntCounter { + pub(crate) struct GlobalAndPerTenantIntCounter { global: IntCounter, per_tenant: IntCounter, } @@ -3583,10 +3594,10 @@ pub(crate) mod tenant_throttling { } pub(crate) struct Metrics { - count_accounted_start: GlobalAndPerTenantIntCounter, - count_accounted_finish: GlobalAndPerTenantIntCounter, - wait_time: GlobalAndPerTenantIntCounter, - count_throttled: GlobalAndPerTenantIntCounter, + pub(super) count_accounted_start: GlobalAndPerTenantIntCounter, + pub(super) count_accounted_finish: GlobalAndPerTenantIntCounter, + pub(super) wait_time: GlobalAndPerTenantIntCounter, + pub(super) count_throttled: GlobalAndPerTenantIntCounter, } static COUNT_ACCOUNTED_START: Lazy = Lazy::new(|| { @@ -3721,26 +3732,6 @@ pub(crate) mod tenant_throttling { } } } - - impl tenant::throttle::Metric for Metrics { - #[inline(always)] - fn accounting_start(&self) { - self.count_accounted_start.inc(); - } - #[inline(always)] - fn accounting_finish(&self) { - self.count_accounted_finish.inc(); - } - #[inline(always)] - fn observe_throttling( - &self, - tenant::throttle::Observation { wait_time }: &tenant::throttle::Observation, - ) { - let val = u64::try_from(wait_time.as_micros()).unwrap(); - self.wait_time.inc_by(val); - self.count_throttled.inc(); - } - } } pub(crate) mod disk_usage_based_eviction { diff --git a/pageserver/src/page_service.rs b/pageserver/src/page_service.rs index f6504bd3b5..b3e18fed99 100644 --- a/pageserver/src/page_service.rs +++ b/pageserver/src/page_service.rs @@ -592,43 +592,21 @@ enum BatchedFeMessage { } impl BatchedFeMessage { - async fn throttle_and_record_start_processing( - &mut self, - cancel: &CancellationToken, - ) -> Result<(), QueryError> { - let (shard, tokens, timers) = match self { - BatchedFeMessage::Exists { shard, timer, .. } - | BatchedFeMessage::Nblocks { shard, timer, .. } - | BatchedFeMessage::DbSize { shard, timer, .. } - | BatchedFeMessage::GetSlruSegment { shard, timer, .. } => { - ( - shard, - // 1 token is probably under-estimating because these - // request handlers typically do several Timeline::get calls. - 1, - itertools::Either::Left(std::iter::once(timer)), - ) + fn observe_execution_start(&mut self, at: Instant) { + match self { + BatchedFeMessage::Exists { timer, .. } + | BatchedFeMessage::Nblocks { timer, .. } + | BatchedFeMessage::DbSize { timer, .. } + | BatchedFeMessage::GetSlruSegment { timer, .. } => { + timer.observe_execution_start(at); } - BatchedFeMessage::GetPage { shard, pages, .. } => ( - shard, - pages.len(), - itertools::Either::Right(pages.iter_mut().map(|p| &mut p.timer)), - ), - BatchedFeMessage::RespondError { .. } => return Ok(()), - }; - let throttled = tokio::select! { - throttled = shard.pagestream_throttle.throttle(tokens) => { throttled } - _ = shard.cancel.cancelled() => { - return Err(QueryError::Shutdown); + BatchedFeMessage::GetPage { pages, .. } => { + for page in pages { + page.timer.observe_execution_start(at); + } } - _ = cancel.cancelled() => { - return Err(QueryError::Shutdown); - } - }; - for timer in timers { - timer.observe_throttle_done_execution_starting(&throttled); + BatchedFeMessage::RespondError { .. } => {} } - Ok(()) } } @@ -720,6 +698,26 @@ impl PageServerHandler { let neon_fe_msg = PagestreamFeMessage::parse(&mut copy_data_bytes.reader(), protocol_version)?; + // TODO: turn in to async closure once available to avoid repeating received_at + async fn record_op_start_and_throttle( + shard: &timeline::handle::Handle, + op: metrics::SmgrQueryType, + received_at: Instant, + ) -> Result { + // It's important to start the smgr op metric recorder as early as possible + // so that the _started counters are incremented before we do + // any serious waiting, e.g., for throttle, batching, or actual request handling. + let mut timer = shard.query_metrics.start_smgr_op(op, received_at); + let now = Instant::now(); + timer.observe_throttle_start(now); + let throttled = tokio::select! { + res = shard.pagestream_throttle.throttle(1, now) => res, + _ = shard.cancel.cancelled() => return Err(QueryError::Shutdown), + }; + timer.observe_throttle_done(throttled); + Ok(timer) + } + let batched_msg = match neon_fe_msg { PagestreamFeMessage::Exists(req) => { let span = tracing::info_span!(parent: parent_span, "handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.hdr.request_lsn); @@ -727,9 +725,12 @@ impl PageServerHandler { .get(tenant_id, timeline_id, ShardSelector::Zero) .instrument(span.clone()) // sets `shard_id` field .await?; - let timer = shard - .query_metrics - .start_smgr_op(metrics::SmgrQueryType::GetRelExists, received_at); + let timer = record_op_start_and_throttle( + &shard, + metrics::SmgrQueryType::GetRelExists, + received_at, + ) + .await?; BatchedFeMessage::Exists { span, timer, @@ -743,9 +744,12 @@ impl PageServerHandler { .get(tenant_id, timeline_id, ShardSelector::Zero) .instrument(span.clone()) // sets `shard_id` field .await?; - let timer = shard - .query_metrics - .start_smgr_op(metrics::SmgrQueryType::GetRelSize, received_at); + let timer = record_op_start_and_throttle( + &shard, + metrics::SmgrQueryType::GetRelSize, + received_at, + ) + .await?; BatchedFeMessage::Nblocks { span, timer, @@ -759,9 +763,12 @@ impl PageServerHandler { .get(tenant_id, timeline_id, ShardSelector::Zero) .instrument(span.clone()) // sets `shard_id` field .await?; - let timer = shard - .query_metrics - .start_smgr_op(metrics::SmgrQueryType::GetDbSize, received_at); + let timer = record_op_start_and_throttle( + &shard, + metrics::SmgrQueryType::GetDbSize, + received_at, + ) + .await?; BatchedFeMessage::DbSize { span, timer, @@ -775,9 +782,12 @@ impl PageServerHandler { .get(tenant_id, timeline_id, ShardSelector::Zero) .instrument(span.clone()) // sets `shard_id` field .await?; - let timer = shard - .query_metrics - .start_smgr_op(metrics::SmgrQueryType::GetSlruSegment, received_at); + let timer = record_op_start_and_throttle( + &shard, + metrics::SmgrQueryType::GetSlruSegment, + received_at, + ) + .await?; BatchedFeMessage::GetSlruSegment { span, timer, @@ -826,12 +836,12 @@ impl PageServerHandler { } }; - // It's important to start the timer before waiting for the LSN - // so that the _started counters are incremented before we do - // any serious waiting, e.g., for LSNs. - let timer = shard - .query_metrics - .start_smgr_op(metrics::SmgrQueryType::GetPageAtLsn, received_at); + let timer = record_op_start_and_throttle( + &shard, + metrics::SmgrQueryType::GetPageAtLsn, + received_at, + ) + .await?; let effective_request_lsn = match Self::wait_or_get_last_lsn( &shard, @@ -937,6 +947,13 @@ impl PageServerHandler { where IO: AsyncRead + AsyncWrite + Send + Sync + Unpin, { + let started_at = Instant::now(); + let batch = { + let mut batch = batch; + batch.observe_execution_start(started_at); + batch + }; + // invoke handler function let (handler_results, span): ( Vec>, @@ -1103,8 +1120,11 @@ impl PageServerHandler { // The timer's underlying metric is used for a storage-internal latency SLO and // we don't want to include latency in it that we can't control. // And as pointed out above, in this case, we don't control the time that flush will take. - let flushing_timer = - timer.map(|timer| timer.observe_smgr_op_completion_and_start_flushing()); + let flushing_timer = timer.map(|mut timer| { + timer + .observe_execution_end_flush_start(Instant::now()) + .expect("we are the first caller") + }); // what we want to do let flush_fut = pgb_writer.flush(); @@ -1258,7 +1278,7 @@ impl PageServerHandler { Ok(msg) => msg, Err(e) => break e, }; - let mut msg = match msg { + let msg = match msg { Some(msg) => msg, None => { debug!("pagestream subprotocol end observed"); @@ -1266,10 +1286,6 @@ impl PageServerHandler { } }; - if let Err(cancelled) = msg.throttle_and_record_start_processing(&self.cancel).await { - break cancelled; - } - let err = self .pagesteam_handle_batched_message(pgb_writer, msg, &cancel, protocol_version, ctx) .await; @@ -1429,15 +1445,12 @@ impl PageServerHandler { return Ok(()); } }; - let mut batch = match batch { + let batch = match batch { Ok(batch) => batch, Err(e) => { return Err(e); } }; - batch - .throttle_and_record_start_processing(&self.cancel) - .await?; self.pagesteam_handle_batched_message( pgb_writer, batch, diff --git a/pageserver/src/tenant.rs b/pageserver/src/tenant.rs index 070593b104..f6d758ad22 100644 --- a/pageserver/src/tenant.rs +++ b/pageserver/src/tenant.rs @@ -365,8 +365,9 @@ pub struct Tenant { /// Throttle applied at the top of [`Timeline::get`]. /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance. - pub(crate) pagestream_throttle: - Arc>, + pub(crate) pagestream_throttle: Arc, + + pub(crate) pagestream_throttle_metrics: Arc, /// An ongoing timeline detach concurrency limiter. /// @@ -1687,6 +1688,7 @@ impl Tenant { TimelineResources { remote_client, pagestream_throttle: self.pagestream_throttle.clone(), + pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(), l0_flush_global_state: self.l0_flush_global_state.clone(), }, LoadTimelineCause::Attach, @@ -3992,6 +3994,9 @@ impl Tenant { Ok(timeline) } + /// [`Tenant::shutdown`] must be called before dropping the returned [`Tenant`] object + /// to ensure proper cleanup of background tasks and metrics. + // // Allow too_many_arguments because a constructor's argument list naturally grows with the // number of attributes in the struct: breaking these out into a builder wouldn't be helpful. #[allow(clippy::too_many_arguments)] @@ -4100,8 +4105,10 @@ impl Tenant { gate: Gate::default(), pagestream_throttle: Arc::new(throttle::Throttle::new( Tenant::get_pagestream_throttle_config(conf, &attached_conf.tenant_conf), - crate::metrics::tenant_throttling::Metrics::new(&tenant_shard_id), )), + pagestream_throttle_metrics: Arc::new( + crate::metrics::tenant_throttling::Pagestream::new(&tenant_shard_id), + ), tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)), ongoing_timeline_detach: std::sync::Mutex::default(), gc_block: Default::default(), @@ -5008,6 +5015,7 @@ impl Tenant { TimelineResources { remote_client: self.build_timeline_remote_client(timeline_id), pagestream_throttle: self.pagestream_throttle.clone(), + pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(), l0_flush_global_state: self.l0_flush_global_state.clone(), } } diff --git a/pageserver/src/tenant/throttle.rs b/pageserver/src/tenant/throttle.rs index 8ab6a0e060..300d779125 100644 --- a/pageserver/src/tenant/throttle.rs +++ b/pageserver/src/tenant/throttle.rs @@ -3,7 +3,7 @@ use std::{ atomic::{AtomicU64, Ordering}, Arc, }, - time::{Duration, Instant}, + time::Instant, }; use arc_swap::ArcSwap; @@ -16,9 +16,8 @@ use utils::leaky_bucket::{LeakyBucketConfig, RateLimiter}; /// To share a throttle among multiple entities, wrap it in an [`Arc`]. /// /// The intial use case for this is tenant-wide throttling of getpage@lsn requests. -pub struct Throttle { +pub struct Throttle { inner: ArcSwap, - metric: M, /// will be turned into [`Stats::count_accounted_start`] count_accounted_start: AtomicU64, /// will be turned into [`Stats::count_accounted_finish`] @@ -36,15 +35,6 @@ pub struct Inner { pub type Config = pageserver_api::models::ThrottleConfig; -pub struct Observation { - pub wait_time: Duration, -} -pub trait Metric { - fn accounting_start(&self); - fn accounting_finish(&self); - fn observe_throttling(&self, observation: &Observation); -} - /// See [`Throttle::reset_stats`]. pub struct Stats { /// Number of requests that started [`Throttle::throttle`] calls. @@ -59,18 +49,14 @@ pub struct Stats { } pub enum ThrottleResult { - NotThrottled { start: Instant }, - Throttled { start: Instant, end: Instant }, + NotThrottled { end: Instant }, + Throttled { end: Instant }, } -impl Throttle -where - M: Metric, -{ - pub fn new(config: Config, metric: M) -> Self { +impl Throttle { + pub fn new(config: Config) -> Self { Self { inner: ArcSwap::new(Arc::new(Self::new_inner(config))), - metric, count_accounted_start: AtomicU64::new(0), count_accounted_finish: AtomicU64::new(0), count_throttled: AtomicU64::new(0), @@ -127,32 +113,27 @@ where self.inner.load().rate_limiter.steady_rps() } - pub async fn throttle(&self, key_count: usize) -> ThrottleResult { + /// `start` must be [`Instant::now`] or earlier. + pub async fn throttle(&self, key_count: usize, start: Instant) -> ThrottleResult { let inner = self.inner.load_full(); // clones the `Inner` Arc - let start = std::time::Instant::now(); - if !inner.enabled { - return ThrottleResult::NotThrottled { start }; + return ThrottleResult::NotThrottled { end: start }; } - self.metric.accounting_start(); self.count_accounted_start.fetch_add(1, Ordering::Relaxed); let did_throttle = inner.rate_limiter.acquire(key_count).await; self.count_accounted_finish.fetch_add(1, Ordering::Relaxed); - self.metric.accounting_finish(); if did_throttle { self.count_throttled.fetch_add(1, Ordering::Relaxed); - let now = Instant::now(); - let wait_time = now - start; + let end = Instant::now(); + let wait_time = end - start; self.sum_throttled_usecs .fetch_add(wait_time.as_micros() as u64, Ordering::Relaxed); - let observation = Observation { wait_time }; - self.metric.observe_throttling(&observation); - ThrottleResult::Throttled { start, end: now } + ThrottleResult::Throttled { end } } else { - ThrottleResult::NotThrottled { start } + ThrottleResult::NotThrottled { end: start } } } } diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs index 741b214a73..4aa6b7a05a 100644 --- a/pageserver/src/tenant/timeline.rs +++ b/pageserver/src/tenant/timeline.rs @@ -208,8 +208,8 @@ fn drop_wlock(rlock: tokio::sync::RwLockWriteGuard<'_, T>) { /// The outward-facing resources required to build a Timeline pub struct TimelineResources { pub remote_client: RemoteTimelineClient, - pub pagestream_throttle: - Arc>, + pub pagestream_throttle: Arc, + pub pagestream_throttle_metrics: Arc, pub l0_flush_global_state: l0_flush::L0FlushGlobalState, } @@ -412,8 +412,7 @@ pub struct Timeline { gc_lock: tokio::sync::Mutex<()>, /// Cloned from [`super::Tenant::pagestream_throttle`] on construction. - pub(crate) pagestream_throttle: - Arc>, + pub(crate) pagestream_throttle: Arc, /// Size estimator for aux file v2 pub(crate) aux_file_size_estimator: AuxFileSizeEstimator, @@ -2310,6 +2309,7 @@ impl Timeline { query_metrics: crate::metrics::SmgrQueryTimePerTimeline::new( &tenant_shard_id, &timeline_id, + resources.pagestream_throttle_metrics, ), directory_metrics: array::from_fn(|_| AtomicU64::new(0)), diff --git a/pageserver/src/tenant/timeline/delete.rs b/pageserver/src/tenant/timeline/delete.rs index ae44af3fad..bdc315d985 100644 --- a/pageserver/src/tenant/timeline/delete.rs +++ b/pageserver/src/tenant/timeline/delete.rs @@ -301,6 +301,7 @@ impl DeleteTimelineFlow { TimelineResources { remote_client, pagestream_throttle: tenant.pagestream_throttle.clone(), + pagestream_throttle_metrics: tenant.pagestream_throttle_metrics.clone(), l0_flush_global_state: tenant.l0_flush_global_state.clone(), }, // Important. We dont pass ancestor above because it can be missing. From aa7323a384dd60b8c176ba6bed6a3c2e5f9bd95e Mon Sep 17 00:00:00 2001 From: John Spray Date: Tue, 14 Jan 2025 15:30:43 +0000 Subject: [PATCH 04/31] storage controller: quality of life improvements for AZ handling (#10379) ## Problem Since https://github.com/neondatabase/neon/pull/9916, the preferred AZ of a tenant is much more impactful, and we would like to make it more visible in tooling. ## Summary of changes - Include AZ in node describe API - Include AZ info in node & tenant outputs in CLI - Add metrics for per-node shard counts, labelled by AZ - Add a CLI for setting preferred AZ on a tenant - Extend AZ-setting API+CLI to handle None for clearing preferred AZ --- control_plane/storcon_cli/src/main.rs | 120 ++++++++++++++++++++-- libs/pageserver_api/src/controller_api.rs | 4 +- storage_controller/src/metrics.rs | 19 ++++ storage_controller/src/node.rs | 1 + storage_controller/src/persistence.rs | 7 +- storage_controller/src/scheduler.rs | 29 +++++- storage_controller/src/service.rs | 9 +- storage_controller/src/tenant_shard.rs | 4 +- 8 files changed, 175 insertions(+), 18 deletions(-) diff --git a/control_plane/storcon_cli/src/main.rs b/control_plane/storcon_cli/src/main.rs index 1653b3c845..9d133e4af1 100644 --- a/control_plane/storcon_cli/src/main.rs +++ b/control_plane/storcon_cli/src/main.rs @@ -1,12 +1,16 @@ use futures::StreamExt; -use std::{str::FromStr, time::Duration}; +use std::{ + collections::{HashMap, HashSet}, + str::FromStr, + time::Duration, +}; use clap::{Parser, Subcommand}; use pageserver_api::{ controller_api::{ AvailabilityZone, NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse, - SafekeeperDescribeResponse, ShardSchedulingPolicy, TenantCreateRequest, - TenantDescribeResponse, TenantPolicyRequest, + SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest, + TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest, }, models::{ EvictionPolicy, EvictionPolicyLayerAccessThreshold, LocationConfigSecondary, @@ -153,6 +157,12 @@ enum Command { #[arg(long)] tenant_id: TenantId, }, + TenantSetPreferredAz { + #[arg(long)] + tenant_id: TenantId, + #[arg(long)] + preferred_az: Option, + }, /// Uncleanly drop a tenant from the storage controller: this doesn't delete anything from pageservers. Appropriate /// if you e.g. used `tenant-warmup` by mistake on a tenant ID that doesn't really exist, or is in some other region. TenantDrop { @@ -402,11 +412,12 @@ async fn main() -> anyhow::Result<()> { resp.sort_by(|a, b| a.listen_http_addr.cmp(&b.listen_http_addr)); let mut table = comfy_table::Table::new(); - table.set_header(["Id", "Hostname", "Scheduling", "Availability"]); + table.set_header(["Id", "Hostname", "AZ", "Scheduling", "Availability"]); for node in resp { table.add_row([ format!("{}", node.id), node.listen_http_addr, + node.availability_zone_id, format!("{:?}", node.scheduling), format!("{:?}", node.availability), ]); @@ -479,6 +490,7 @@ async fn main() -> anyhow::Result<()> { let mut table = comfy_table::Table::new(); table.set_header([ "TenantId", + "Preferred AZ", "ShardCount", "StripeSize", "Placement", @@ -488,6 +500,11 @@ async fn main() -> anyhow::Result<()> { let shard_zero = tenant.shards.into_iter().next().unwrap(); table.add_row([ format!("{}", tenant.tenant_id), + shard_zero + .preferred_az_id + .as_ref() + .cloned() + .unwrap_or("".to_string()), format!("{}", shard_zero.tenant_shard_id.shard_count.literal()), format!("{:?}", tenant.stripe_size), format!("{:?}", tenant.policy), @@ -614,6 +631,19 @@ async fn main() -> anyhow::Result<()> { None, ) .await?; + + let nodes = storcon_client + .dispatch::<(), Vec>( + Method::GET, + "control/v1/node".to_string(), + None, + ) + .await?; + let nodes = nodes + .into_iter() + .map(|n| (n.id, n)) + .collect::>(); + println!("Tenant {tenant_id}"); let mut table = comfy_table::Table::new(); table.add_row(["Policy", &format!("{:?}", policy)]); @@ -622,7 +652,14 @@ async fn main() -> anyhow::Result<()> { println!("{table}"); println!("Shards:"); let mut table = comfy_table::Table::new(); - table.set_header(["Shard", "Attached", "Secondary", "Last error", "status"]); + table.set_header([ + "Shard", + "Attached", + "Attached AZ", + "Secondary", + "Last error", + "status", + ]); for shard in shards { let secondary = shard .node_secondary @@ -645,11 +682,18 @@ async fn main() -> anyhow::Result<()> { } let status = status_parts.join(","); + let attached_node = shard + .node_attached + .as_ref() + .map(|id| nodes.get(id).expect("Shard references nonexistent node")); + table.add_row([ format!("{}", shard.tenant_shard_id), - shard - .node_attached - .map(|n| format!("{}", n)) + attached_node + .map(|n| format!("{} ({})", n.listen_http_addr, n.id)) + .unwrap_or(String::new()), + attached_node + .map(|n| n.availability_zone_id.clone()) .unwrap_or(String::new()), secondary, shard.last_error, @@ -658,6 +702,66 @@ async fn main() -> anyhow::Result<()> { } println!("{table}"); } + Command::TenantSetPreferredAz { + tenant_id, + preferred_az, + } => { + // First learn about the tenant's shards + let describe_response = storcon_client + .dispatch::<(), TenantDescribeResponse>( + Method::GET, + format!("control/v1/tenant/{tenant_id}"), + None, + ) + .await?; + + // Learn about nodes to validate the AZ ID + let nodes = storcon_client + .dispatch::<(), Vec>( + Method::GET, + "control/v1/node".to_string(), + None, + ) + .await?; + + if let Some(preferred_az) = &preferred_az { + let azs = nodes + .into_iter() + .map(|n| (n.availability_zone_id)) + .collect::>(); + if !azs.contains(preferred_az) { + anyhow::bail!( + "AZ {} not found on any node: known AZs are: {:?}", + preferred_az, + azs + ); + } + } else { + // Make it obvious to the user that since they've omitted an AZ, we're clearing it + eprintln!("Clearing preferred AZ for tenant {}", tenant_id); + } + + // Construct a request that modifies all the tenant's shards + let req = ShardsPreferredAzsRequest { + preferred_az_ids: describe_response + .shards + .into_iter() + .map(|s| { + ( + s.tenant_shard_id, + preferred_az.clone().map(AvailabilityZone), + ) + }) + .collect(), + }; + storcon_client + .dispatch::( + Method::PUT, + "control/v1/preferred_azs".to_string(), + Some(req), + ) + .await?; + } Command::TenantWarmup { tenant_id } => { let describe_response = storcon_client .dispatch::<(), TenantDescribeResponse>( diff --git a/libs/pageserver_api/src/controller_api.rs b/libs/pageserver_api/src/controller_api.rs index f3aefc6df9..f3880cb766 100644 --- a/libs/pageserver_api/src/controller_api.rs +++ b/libs/pageserver_api/src/controller_api.rs @@ -87,7 +87,7 @@ impl Display for AvailabilityZone { #[derive(Serialize, Deserialize)] pub struct ShardsPreferredAzsRequest { #[serde(flatten)] - pub preferred_az_ids: HashMap, + pub preferred_az_ids: HashMap>, } #[derive(Serialize, Deserialize)] @@ -144,6 +144,8 @@ pub struct NodeDescribeResponse { pub availability: NodeAvailabilityWrapper, pub scheduling: NodeSchedulingPolicy, + pub availability_zone_id: String, + pub listen_http_addr: String, pub listen_http_port: u16, diff --git a/storage_controller/src/metrics.rs b/storage_controller/src/metrics.rs index 6d5885eba6..4164e3dc2b 100644 --- a/storage_controller/src/metrics.rs +++ b/storage_controller/src/metrics.rs @@ -53,6 +53,16 @@ pub(crate) struct StorageControllerMetricGroup { /// How many shards are not scheduled into their preferred AZ pub(crate) storage_controller_schedule_az_violation: measured::Gauge, + /// How many shard locations (secondary or attached) on each node + pub(crate) storage_controller_node_shards: measured::GaugeVec, + + /// How many _attached_ shard locations on each node + pub(crate) storage_controller_node_attached_shards: measured::GaugeVec, + + /// How many _home_ shard locations on each node (i.e. the node's AZ matches the shard's + /// preferred AZ) + pub(crate) storage_controller_node_home_shards: measured::GaugeVec, + /// How many shards would like to reconcile but were blocked by concurrency limits pub(crate) storage_controller_pending_reconciles: measured::Gauge, @@ -132,6 +142,15 @@ impl Default for StorageControllerMetrics { } } +#[derive(measured::LabelGroup, Clone)] +#[label(set = NodeLabelGroupSet)] +pub(crate) struct NodeLabelGroup<'a> { + #[label(dynamic_with = lasso::ThreadedRodeo, default)] + pub(crate) az: &'a str, + #[label(dynamic_with = lasso::ThreadedRodeo, default)] + pub(crate) node_id: &'a str, +} + #[derive(measured::LabelGroup)] #[label(set = ReconcileCompleteLabelGroupSet)] pub(crate) struct ReconcileCompleteLabelGroup { diff --git a/storage_controller/src/node.rs b/storage_controller/src/node.rs index 4cc9b0070d..f5c2d329e0 100644 --- a/storage_controller/src/node.rs +++ b/storage_controller/src/node.rs @@ -299,6 +299,7 @@ impl Node { id: self.id, availability: self.availability.clone().into(), scheduling: self.scheduling, + availability_zone_id: self.availability_zone_id.0.clone(), listen_http_addr: self.listen_http_addr.clone(), listen_http_port: self.listen_http_port, listen_pg_addr: self.listen_pg_addr.clone(), diff --git a/storage_controller/src/persistence.rs b/storage_controller/src/persistence.rs index beb014f0a8..eb0bfc879e 100644 --- a/storage_controller/src/persistence.rs +++ b/storage_controller/src/persistence.rs @@ -708,10 +708,11 @@ impl Persistence { Ok(()) } + /// Note that passing None for a shard clears the preferred AZ (rather than leaving it unmodified) pub(crate) async fn set_tenant_shard_preferred_azs( &self, - preferred_azs: Vec<(TenantShardId, AvailabilityZone)>, - ) -> DatabaseResult> { + preferred_azs: Vec<(TenantShardId, Option)>, + ) -> DatabaseResult)>> { use crate::schema::tenant_shards::dsl::*; self.with_measured_conn(DatabaseOperation::SetPreferredAzs, move |conn| { @@ -722,7 +723,7 @@ impl Persistence { .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string())) .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32)) .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32)) - .set(preferred_az_id.eq(preferred_az.0.clone())) + .set(preferred_az_id.eq(preferred_az.as_ref().map(|az| az.0.clone()))) .execute(conn)?; if updated == 1 { diff --git a/storage_controller/src/scheduler.rs b/storage_controller/src/scheduler.rs index 04a594dcac..f5cab9dd57 100644 --- a/storage_controller/src/scheduler.rs +++ b/storage_controller/src/scheduler.rs @@ -1,4 +1,4 @@ -use crate::{node::Node, tenant_shard::TenantShard}; +use crate::{metrics::NodeLabelGroup, node::Node, tenant_shard::TenantShard}; use itertools::Itertools; use pageserver_api::{controller_api::AvailabilityZone, models::PageserverUtilization}; use serde::Serialize; @@ -872,6 +872,33 @@ impl Scheduler { pub(crate) fn get_node_attached_shard_count(&self, node_id: NodeId) -> usize { self.nodes.get(&node_id).unwrap().attached_shard_count } + + /// Some metrics that we only calculate periodically: this is simpler than + /// rigorously updating them on every change. + pub(crate) fn update_metrics(&self) { + for (node_id, node) in &self.nodes { + let node_id_str = format!("{}", node_id); + let label_group = NodeLabelGroup { + az: &node.az.0, + node_id: &node_id_str, + }; + + crate::metrics::METRICS_REGISTRY + .metrics_group + .storage_controller_node_shards + .set(label_group.clone(), node.shard_count as i64); + + crate::metrics::METRICS_REGISTRY + .metrics_group + .storage_controller_node_attached_shards + .set(label_group.clone(), node.attached_shard_count as i64); + + crate::metrics::METRICS_REGISTRY + .metrics_group + .storage_controller_node_home_shards + .set(label_group.clone(), node.home_shard_count as i64); + } + } } #[cfg(test)] diff --git a/storage_controller/src/service.rs b/storage_controller/src/service.rs index dadcc44cfb..cbb9103880 100644 --- a/storage_controller/src/service.rs +++ b/storage_controller/src/service.rs @@ -2517,7 +2517,7 @@ impl Service { .map(|t| { ( t.get_tenant_shard_id().expect("Corrupt shard in database"), - load_in_az.clone(), + Some(load_in_az.clone()), ) }) .collect(), @@ -6390,7 +6390,7 @@ impl Service { /// available. A return value of 0 indicates that everything is fully reconciled already. fn reconcile_all(&self) -> usize { let mut locked = self.inner.write().unwrap(); - let (nodes, tenants, _scheduler) = locked.parts_mut(); + let (nodes, tenants, scheduler) = locked.parts_mut(); let pageservers = nodes.clone(); // This function is an efficient place to update lazy statistics, since we are walking @@ -6451,6 +6451,9 @@ impl Service { } } + // Some metrics are calculated from SchedulerNode state, update these periodically + scheduler.update_metrics(); + // Process any deferred tenant drops for (tenant_id, guard) in drop_detached_tenants { self.maybe_drop_tenant(tenant_id, &mut locked, &guard); @@ -6509,7 +6512,7 @@ impl Service { // Shard was dropped between planning and execution; continue; }; - tracing::info!("Applying optimization: {optimization:?}"); + tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}"); if shard.apply_optimization(scheduler, optimization) { optimizations_applied += 1; if self.maybe_reconcile_shard(shard, nodes).is_some() { diff --git a/storage_controller/src/tenant_shard.rs b/storage_controller/src/tenant_shard.rs index 2ba2a57eba..79ed628c25 100644 --- a/storage_controller/src/tenant_shard.rs +++ b/storage_controller/src/tenant_shard.rs @@ -1708,8 +1708,8 @@ impl TenantShard { self.intent.preferred_az_id.as_ref() } - pub(crate) fn set_preferred_az(&mut self, preferred_az_id: AvailabilityZone) { - self.intent.preferred_az_id = Some(preferred_az_id); + pub(crate) fn set_preferred_az(&mut self, preferred_az_id: Option) { + self.intent.preferred_az_id = preferred_az_id; } /// Returns all the nodes to which this tenant shard is attached according to the From ffaa52ff5d3f36becaf70be2b6053c7423deba61 Mon Sep 17 00:00:00 2001 From: Erik Grinaker Date: Tue, 14 Jan 2025 17:31:59 +0100 Subject: [PATCH 05/31] pageserver: reorder upload queue when possible (#10218) ## Problem The upload queue currently sees significant head-of-line blocking. For example, index uploads act as upload barriers, and for every layer flush we schedule a layer and index upload, which effectively serializes layer uploads. Resolves #10096. ## Summary of changes Allow upload queue operations to bypass the queue if they don't conflict with preceding operations, increasing parallelism. NB: the upload queue currently schedules an explicit barrier after every layer flush as well (see #8550). This must be removed to enable parallelism. This will require a better mechanism for compaction backpressure, see e.g. #8390 or #5415. --- Cargo.lock | 1 + pageserver/Cargo.toml | 5 + pageserver/benches/upload_queue.rs | 86 ++ pageserver/src/tenant/metadata.rs | 1 - .../src/tenant/remote_timeline_client.rs | 101 +- .../tenant/remote_timeline_client/index.rs | 2 +- pageserver/src/tenant/storage_layer/layer.rs | 2 +- pageserver/src/tenant/upload_queue.rs | 957 +++++++++++++++++- 8 files changed, 1029 insertions(+), 126 deletions(-) create mode 100644 pageserver/benches/upload_queue.rs diff --git a/Cargo.lock b/Cargo.lock index 08453120c7..1e29f4fc08 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -4044,6 +4044,7 @@ dependencies = [ "postgres_connection", "postgres_ffi", "postgres_initdb", + "pprof", "pq_proto", "procfs", "rand 0.8.5", diff --git a/pageserver/Cargo.toml b/pageserver/Cargo.toml index 140b287ccc..8547746d94 100644 --- a/pageserver/Cargo.toml +++ b/pageserver/Cargo.toml @@ -44,6 +44,7 @@ postgres_backend.workspace = true postgres-protocol.workspace = true postgres-types.workspace = true postgres_initdb.workspace = true +pprof.workspace = true rand.workspace = true range-set-blaze = { version = "0.1.16", features = ["alloc"] } regex.workspace = true @@ -108,3 +109,7 @@ harness = false [[bench]] name = "bench_ingest" harness = false + +[[bench]] +name = "upload_queue" +harness = false diff --git a/pageserver/benches/upload_queue.rs b/pageserver/benches/upload_queue.rs new file mode 100644 index 0000000000..528b3d5490 --- /dev/null +++ b/pageserver/benches/upload_queue.rs @@ -0,0 +1,86 @@ +//! Upload queue benchmarks. + +use std::str::FromStr as _; +use std::sync::atomic::AtomicU32; +use std::sync::Arc; + +use criterion::{criterion_group, criterion_main, Bencher, Criterion}; +use pageserver::tenant::metadata::TimelineMetadata; +use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata; +use pageserver::tenant::storage_layer::LayerName; +use pageserver::tenant::upload_queue::{Delete, UploadOp, UploadQueue, UploadTask}; +use pageserver::tenant::IndexPart; +use pprof::criterion::{Output, PProfProfiler}; +use utils::generation::Generation; +use utils::shard::{ShardCount, ShardIndex, ShardNumber}; + +// Register benchmarks with Criterion. +criterion_group!( + name = benches; + config = Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None))); + targets = bench_upload_queue_next_ready, +); +criterion_main!(benches); + +/// Benchmarks the cost of UploadQueue::next_ready() with the given number of in-progress tasks +/// (which is equivalent to tasks ahead of it in the queue). This has linear cost, and the upload +/// queue as a whole is thus quadratic. +/// +/// UploadOp::UploadLayer requires an entire tenant and timeline to construct, so we just test +/// Delete and UploadMetadata instead. This is incidentally the most expensive case. +fn bench_upload_queue_next_ready(c: &mut Criterion) { + let mut g = c.benchmark_group("upload_queue_next_ready"); + for inprogress in [0, 1, 10, 100, 1_000, 10_000, 100_000, 1_000_000] { + g.bench_function(format!("inprogress={inprogress}"), |b| { + run_bench(b, inprogress).unwrap() + }); + } + + fn run_bench(b: &mut Bencher, inprogress: usize) -> anyhow::Result<()> { + // Construct two layers. layer0 is in the indexes, layer1 will be deleted. + let layer0 = LayerName::from_str("000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51").expect("invalid name"); + let layer1 = LayerName::from_str("100000000000000000000000000000000001-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51").expect("invalid name"); + + let metadata = LayerFileMetadata { + shard: ShardIndex::new(ShardNumber(1), ShardCount(2)), + generation: Generation::Valid(1), + file_size: 0, + }; + + // Construct the (initial and uploaded) index with layer0. + let mut index = IndexPart::empty(TimelineMetadata::example()); + index.layer_metadata.insert(layer0, metadata.clone()); + + // Construct the queue. + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&index)?; + + // Populate inprogress_tasks with a bunch of layer1 deletions. + let delete = UploadOp::Delete(Delete { + layers: vec![(layer1, metadata)], + }); + + for task_id in 0..(inprogress as u64) { + queue.inprogress_tasks.insert( + task_id, + Arc::new(UploadTask { + task_id, + retries: AtomicU32::new(0), + op: delete.clone(), + }), + ); + } + + // Benchmark index upload scheduling. + let index_upload = UploadOp::UploadMetadata { + uploaded: Box::new(index), + }; + + b.iter(|| { + queue.queued_operations.push_front(index_upload.clone()); + assert!(queue.next_ready().is_some()); + }); + + Ok(()) + } +} diff --git a/pageserver/src/tenant/metadata.rs b/pageserver/src/tenant/metadata.rs index 24440d4b35..d281eb305f 100644 --- a/pageserver/src/tenant/metadata.rs +++ b/pageserver/src/tenant/metadata.rs @@ -320,7 +320,6 @@ impl TimelineMetadata { // Checksums make it awkward to build a valid instance by hand. This helper // provides a TimelineMetadata with a valid checksum in its header. - #[cfg(test)] pub fn example() -> Self { let instance = Self::new( "0/16960E8".parse::().unwrap(), diff --git a/pageserver/src/tenant/remote_timeline_client.rs b/pageserver/src/tenant/remote_timeline_client.rs index 813111245d..75e8da496d 100644 --- a/pageserver/src/tenant/remote_timeline_client.rs +++ b/pageserver/src/tenant/remote_timeline_client.rs @@ -63,22 +63,18 @@ //! The contract between client and its user is that the user is responsible of //! scheduling operations in an order that keeps the remote consistent as //! described above. +//! //! From the user's perspective, the operations are executed sequentially. //! Internally, the client knows which operations can be performed in parallel, //! and which operations act like a "barrier" that require preceding operations //! to finish. The calling code just needs to call the schedule-functions in the //! correct order, and the client will parallelize the operations in a way that -//! is safe. -//! -//! The caller should be careful with deletion, though. They should not delete -//! local files that have been scheduled for upload but not yet finished uploading. -//! Otherwise the upload will fail. To wait for an upload to finish, use -//! the 'wait_completion' function (more on that later.) +//! is safe. For more details, see `UploadOp::can_bypass`. //! //! All of this relies on the following invariants: //! //! - We rely on read-after write consistency in the remote storage. -//! - Layer files are immutable +//! - Layer files are immutable. //! //! NB: Pageserver assumes that it has exclusive write access to the tenant in remote //! storage. Different tenants can be attached to different pageservers, but if the @@ -1855,57 +1851,17 @@ impl RemoteTimelineClient { Ok(()) } - /// /// Pick next tasks from the queue, and start as many of them as possible without violating /// the ordering constraints. /// - /// The caller needs to already hold the `upload_queue` lock. + /// TODO: consider limiting the number of in-progress tasks, beyond what remote_storage does. + /// This can launch an unbounded number of queued tasks. `UploadQueue::next_ready()` also has + /// worst-case quadratic cost in the number of tasks, and may struggle beyond 10,000 tasks. fn launch_queued_tasks(self: &Arc, upload_queue: &mut UploadQueueInitialized) { - while let Some(next_op) = upload_queue.queued_operations.front() { - // Can we run this task now? - let can_run_now = match next_op { - UploadOp::UploadLayer(..) => { - // Can always be scheduled. - true - } - UploadOp::UploadMetadata { .. } => { - // These can only be performed after all the preceding operations - // have finished. - upload_queue.inprogress_tasks.is_empty() - } - UploadOp::Delete(..) => { - // Wait for preceding uploads to finish. Concurrent deletions are OK, though. - upload_queue.num_inprogress_deletions == upload_queue.inprogress_tasks.len() - } + while let Some(mut next_op) = upload_queue.next_ready() { + debug!("starting op: {next_op}"); - UploadOp::Barrier(_) | UploadOp::Shutdown => { - upload_queue.inprogress_tasks.is_empty() - } - }; - - // If we cannot launch this task, don't look any further. - // - // In some cases, we could let some non-frontmost tasks to "jump the queue" and launch - // them now, but we don't try to do that currently. For example, if the frontmost task - // is an index-file upload that cannot proceed until preceding uploads have finished, we - // could still start layer uploads that were scheduled later. - if !can_run_now { - break; - } - - if let UploadOp::Shutdown = next_op { - // leave the op in the queue but do not start more tasks; it will be dropped when - // the stop is called. - upload_queue.shutdown_ready.close(); - break; - } - - // We can launch this task. Remove it from the queue first. - let mut next_op = upload_queue.queued_operations.pop_front().unwrap(); - - debug!("starting op: {}", next_op); - - // Update the counters and prepare + // Prepare upload. match &mut next_op { UploadOp::UploadLayer(layer, meta, mode) => { if upload_queue @@ -1916,18 +1872,14 @@ impl RemoteTimelineClient { } else { *mode = Some(OpType::MayReorder) } - upload_queue.num_inprogress_layer_uploads += 1; - } - UploadOp::UploadMetadata { .. } => { - upload_queue.num_inprogress_metadata_uploads += 1; } + UploadOp::UploadMetadata { .. } => {} UploadOp::Delete(Delete { layers }) => { for (name, meta) in layers { upload_queue .recently_deleted .insert((name.clone(), meta.generation)); } - upload_queue.num_inprogress_deletions += 1; } UploadOp::Barrier(sender) => { sender.send_replace(()); @@ -2027,6 +1979,8 @@ impl RemoteTimelineClient { let upload_result: anyhow::Result<()> = match &task.op { UploadOp::UploadLayer(ref layer, ref layer_metadata, mode) => { + // TODO: check if this mechanism can be removed now that can_bypass() performs + // conflict checks during scheduling. if let Some(OpType::FlushDeletion) = mode { if self.config.read().unwrap().block_deletions { // Of course, this is not efficient... but usually the queue should be empty. @@ -2249,13 +2203,8 @@ impl RemoteTimelineClient { upload_queue.inprogress_tasks.remove(&task.task_id); let lsn_update = match task.op { - UploadOp::UploadLayer(_, _, _) => { - upload_queue.num_inprogress_layer_uploads -= 1; - None - } + UploadOp::UploadLayer(_, _, _) => None, UploadOp::UploadMetadata { ref uploaded } => { - upload_queue.num_inprogress_metadata_uploads -= 1; - // the task id is reused as a monotonicity check for storing the "clean" // IndexPart. let last_updater = upload_queue.clean.1; @@ -2289,10 +2238,7 @@ impl RemoteTimelineClient { None } } - UploadOp::Delete(_) => { - upload_queue.num_inprogress_deletions -= 1; - None - } + UploadOp::Delete(_) => None, UploadOp::Barrier(..) | UploadOp::Shutdown => unreachable!(), }; @@ -2416,9 +2362,6 @@ impl RemoteTimelineClient { visible_remote_consistent_lsn: initialized .visible_remote_consistent_lsn .clone(), - num_inprogress_layer_uploads: 0, - num_inprogress_metadata_uploads: 0, - num_inprogress_deletions: 0, inprogress_tasks: HashMap::default(), queued_operations: VecDeque::default(), #[cfg(feature = "testing")] @@ -2445,14 +2388,6 @@ impl RemoteTimelineClient { } }; - // consistency check - assert_eq!( - qi.num_inprogress_layer_uploads - + qi.num_inprogress_metadata_uploads - + qi.num_inprogress_deletions, - qi.inprogress_tasks.len() - ); - // We don't need to do anything here for in-progress tasks. They will finish // on their own, decrement the unfinished-task counter themselves, and observe // that the queue is Stopped. @@ -2899,8 +2834,8 @@ mod tests { let mut guard = client.upload_queue.lock().unwrap(); let upload_queue = guard.initialized_mut().unwrap(); assert!(upload_queue.queued_operations.is_empty()); - assert!(upload_queue.inprogress_tasks.len() == 2); - assert!(upload_queue.num_inprogress_layer_uploads == 2); + assert_eq!(upload_queue.inprogress_tasks.len(), 2); + assert_eq!(upload_queue.num_inprogress_layer_uploads(), 2); // also check that `latest_file_changes` was updated assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 2); @@ -2970,8 +2905,8 @@ mod tests { // Deletion schedules upload of the index file, and the file deletion itself assert_eq!(upload_queue.queued_operations.len(), 2); assert_eq!(upload_queue.inprogress_tasks.len(), 1); - assert_eq!(upload_queue.num_inprogress_layer_uploads, 1); - assert_eq!(upload_queue.num_inprogress_deletions, 0); + assert_eq!(upload_queue.num_inprogress_layer_uploads(), 1); + assert_eq!(upload_queue.num_inprogress_deletions(), 0); assert_eq!( upload_queue.latest_files_changes_since_metadata_upload_scheduled, 0 diff --git a/pageserver/src/tenant/remote_timeline_client/index.rs b/pageserver/src/tenant/remote_timeline_client/index.rs index 51f093cb87..244be5bbb7 100644 --- a/pageserver/src/tenant/remote_timeline_client/index.rs +++ b/pageserver/src/tenant/remote_timeline_client/index.rs @@ -104,7 +104,7 @@ impl IndexPart { pub const FILE_NAME: &'static str = "index_part.json"; - pub(crate) fn empty(metadata: TimelineMetadata) -> Self { + pub fn empty(metadata: TimelineMetadata) -> Self { IndexPart { version: Self::LATEST_VERSION, layer_metadata: Default::default(), diff --git a/pageserver/src/tenant/storage_layer/layer.rs b/pageserver/src/tenant/storage_layer/layer.rs index 8933e8ceb1..2b06c88e8b 100644 --- a/pageserver/src/tenant/storage_layer/layer.rs +++ b/pageserver/src/tenant/storage_layer/layer.rs @@ -1812,7 +1812,7 @@ enum LayerKind { /// Guard for forcing a layer be resident while it exists. #[derive(Clone)] -pub(crate) struct ResidentLayer { +pub struct ResidentLayer { owner: Layer, downloaded: Arc, } diff --git a/pageserver/src/tenant/upload_queue.rs b/pageserver/src/tenant/upload_queue.rs index ef3aa759f3..bd524e8153 100644 --- a/pageserver/src/tenant/upload_queue.rs +++ b/pageserver/src/tenant/upload_queue.rs @@ -1,28 +1,33 @@ +use std::collections::{HashMap, HashSet, VecDeque}; +use std::fmt::Debug; +use std::sync::atomic::AtomicU32; +use std::sync::Arc; + +use super::remote_timeline_client::is_same_remote_layer_path; +use super::storage_layer::AsLayerDesc as _; use super::storage_layer::LayerName; use super::storage_layer::ResidentLayer; use crate::tenant::metadata::TimelineMetadata; use crate::tenant::remote_timeline_client::index::IndexPart; use crate::tenant::remote_timeline_client::index::LayerFileMetadata; -use std::collections::HashSet; -use std::collections::{HashMap, VecDeque}; -use std::fmt::Debug; +use utils::generation::Generation; +use utils::lsn::{AtomicLsn, Lsn}; use chrono::NaiveDateTime; -use std::sync::Arc; +use once_cell::sync::Lazy; use tracing::info; -use utils::lsn::AtomicLsn; -use std::sync::atomic::AtomicU32; -use utils::lsn::Lsn; - -use utils::generation::Generation; +/// Kill switch for upload queue reordering in case it causes problems. +/// TODO: remove this once we have confidence in it. +static DISABLE_UPLOAD_QUEUE_REORDERING: Lazy = + Lazy::new(|| std::env::var("DISABLE_UPLOAD_QUEUE_REORDERING").as_deref() == Ok("true")); // clippy warns that Uninitialized is much smaller than Initialized, which wastes // memory for Uninitialized variants. Doesn't matter in practice, there are not // that many upload queues in a running pageserver, and most of them are initialized // anyway. #[allow(clippy::large_enum_variant)] -pub(super) enum UploadQueue { +pub enum UploadQueue { Uninitialized, Initialized(UploadQueueInitialized), Stopped(UploadQueueStopped), @@ -39,13 +44,13 @@ impl UploadQueue { } #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)] -pub(crate) enum OpType { +pub enum OpType { MayReorder, FlushDeletion, } /// This keeps track of queued and in-progress tasks. -pub(crate) struct UploadQueueInitialized { +pub struct UploadQueueInitialized { /// Counter to assign task IDs pub(crate) task_counter: u64, @@ -70,21 +75,16 @@ pub(crate) struct UploadQueueInitialized { /// we skip validation) pub(crate) visible_remote_consistent_lsn: Arc, - // Breakdown of different kinds of tasks currently in-progress - pub(crate) num_inprogress_layer_uploads: usize, - pub(crate) num_inprogress_metadata_uploads: usize, - pub(crate) num_inprogress_deletions: usize, - /// Tasks that are currently in-progress. In-progress means that a tokio Task /// has been launched for it. An in-progress task can be busy uploading, but it can /// also be waiting on the `concurrency_limiter` Semaphore in S3Bucket, or it can /// be waiting for retry in `exponential_backoff`. - pub(crate) inprogress_tasks: HashMap>, + pub inprogress_tasks: HashMap>, /// Queued operations that have not been launched yet. They might depend on previous /// tasks to finish. For example, metadata upload cannot be performed before all /// preceding layer file uploads have completed. - pub(crate) queued_operations: VecDeque, + pub queued_operations: VecDeque, /// Files which have been unlinked but not yet had scheduled a deletion for. Only kept around /// for error logging. @@ -122,6 +122,129 @@ impl UploadQueueInitialized { let lsn = self.clean.0.metadata.disk_consistent_lsn(); self.clean.1.map(|_| lsn) } + + /// Returns and removes the next ready operation from the queue, if any. This isn't necessarily + /// the first operation in the queue, to avoid head-of-line blocking -- an operation can jump + /// the queue if it doesn't conflict with operations ahead of it. + /// + /// None may be returned even if the queue isn't empty, if no operations are ready yet. + pub fn next_ready(&mut self) -> Option { + // NB: this is quadratic, but queues are expected to be small. + for (i, candidate) in self.queued_operations.iter().enumerate() { + // If this candidate is ready, go for it. Otherwise, try the next one. + if self.is_ready(i) { + // Shutdown operations are left at the head of the queue, to prevent further + // operations from starting. Signal that we're ready to shut down. + if matches!(candidate, UploadOp::Shutdown) { + assert!(self.inprogress_tasks.is_empty(), "shutdown with tasks"); + assert_eq!(i, 0, "shutdown not at head of queue"); + self.shutdown_ready.close(); + return None; + } + + return self.queued_operations.remove(i); + } + + // Nothing can bypass a barrier or shutdown. If it wasn't scheduled above, give up. + if matches!(candidate, UploadOp::Barrier(_) | UploadOp::Shutdown) { + return None; + } + + // If upload queue reordering is disabled, bail out after the first operation. + if *DISABLE_UPLOAD_QUEUE_REORDERING { + return None; + } + } + None + } + + /// Returns true if the queued operation at the given position is ready to be uploaded, i.e. if + /// it doesn't conflict with any in-progress or queued operations ahead of it. Operations are + /// allowed to skip the queue when it's safe to do so, to increase parallelism. + /// + /// The position must be valid for the queue size. + fn is_ready(&self, pos: usize) -> bool { + let candidate = self.queued_operations.get(pos).expect("invalid position"); + self + // Look at in-progress operations, in random order. + .inprogress_tasks + .values() + .map(|task| &task.op) + // Then queued operations ahead of the candidate, front-to-back. + .chain(self.queued_operations.iter().take(pos)) + // Keep track of the active index ahead of each operation. This is used to ensure that + // an upload doesn't skip the queue too far, such that it modifies a layer that's + // referenced by an active index. + // + // It's okay that in-progress operations are emitted in random order above, since at + // most one of them can be an index upload (enforced by can_bypass). + .scan(&self.clean.0, |next_active_index, op| { + let active_index = *next_active_index; + if let UploadOp::UploadMetadata { ref uploaded } = op { + *next_active_index = uploaded; // stash index for next operation after this + } + Some((op, active_index)) + }) + // Check if the candidate can bypass all of them. + .all(|(op, active_index)| candidate.can_bypass(op, active_index)) + } + + /// Returns the number of in-progress deletion operations. + #[cfg(test)] + pub(crate) fn num_inprogress_deletions(&self) -> usize { + self.inprogress_tasks + .iter() + .filter(|(_, t)| matches!(t.op, UploadOp::Delete(_))) + .count() + } + + /// Returns the number of in-progress layer uploads. + #[cfg(test)] + pub(crate) fn num_inprogress_layer_uploads(&self) -> usize { + self.inprogress_tasks + .iter() + .filter(|(_, t)| matches!(t.op, UploadOp::UploadLayer(_, _, _))) + .count() + } + + /// Test helper that schedules all ready operations into inprogress_tasks, and returns + /// references to them. + /// + /// TODO: the corresponding production logic should be moved from RemoteTimelineClient into + /// UploadQueue, so we can use the same code path. + #[cfg(test)] + fn schedule_ready(&mut self) -> Vec> { + let mut tasks = Vec::new(); + // NB: schedule operations one by one, to handle conflicts with inprogress_tasks. + while let Some(op) = self.next_ready() { + self.task_counter += 1; + let task = Arc::new(UploadTask { + task_id: self.task_counter, + op, + retries: 0.into(), + }); + self.inprogress_tasks.insert(task.task_id, task.clone()); + tasks.push(task); + } + tasks + } + + /// Test helper that marks an operation as completed, removing it from inprogress_tasks. + /// + /// TODO: the corresponding production logic should be moved from RemoteTimelineClient into + /// UploadQueue, so we can use the same code path. + #[cfg(test)] + fn complete(&mut self, task_id: u64) { + let Some(task) = self.inprogress_tasks.remove(&task_id) else { + return; + }; + // Update the clean index on uploads. + if let UploadOp::UploadMetadata { ref uploaded } = task.op { + if task.task_id > self.clean.1.unwrap_or_default() { + self.clean = (*uploaded.clone(), Some(task.task_id)); + } + } + } } #[derive(Clone, Copy)] @@ -131,12 +254,12 @@ pub(super) enum SetDeletedFlagProgress { Successful(NaiveDateTime), } -pub(super) struct UploadQueueStoppedDeletable { +pub struct UploadQueueStoppedDeletable { pub(super) upload_queue_for_deletion: UploadQueueInitialized, pub(super) deleted_at: SetDeletedFlagProgress, } -pub(super) enum UploadQueueStopped { +pub enum UploadQueueStopped { Deletable(UploadQueueStoppedDeletable), Uninitialized, } @@ -163,7 +286,7 @@ impl NotInitialized { } impl UploadQueue { - pub(crate) fn initialize_empty_remote( + pub fn initialize_empty_remote( &mut self, metadata: &TimelineMetadata, ) -> anyhow::Result<&mut UploadQueueInitialized> { @@ -185,9 +308,6 @@ impl UploadQueue { visible_remote_consistent_lsn: Arc::new(AtomicLsn::new(0)), // what follows are boring default initializations task_counter: 0, - num_inprogress_layer_uploads: 0, - num_inprogress_metadata_uploads: 0, - num_inprogress_deletions: 0, inprogress_tasks: HashMap::new(), queued_operations: VecDeque::new(), #[cfg(feature = "testing")] @@ -202,7 +322,7 @@ impl UploadQueue { Ok(self.initialized_mut().expect("we just set it")) } - pub(crate) fn initialize_with_current_remote_index_part( + pub fn initialize_with_current_remote_index_part( &mut self, index_part: &IndexPart, ) -> anyhow::Result<&mut UploadQueueInitialized> { @@ -227,9 +347,6 @@ impl UploadQueue { ), // what follows are boring default initializations task_counter: 0, - num_inprogress_layer_uploads: 0, - num_inprogress_metadata_uploads: 0, - num_inprogress_deletions: 0, inprogress_tasks: HashMap::new(), queued_operations: VecDeque::new(), #[cfg(feature = "testing")] @@ -244,9 +361,7 @@ impl UploadQueue { Ok(self.initialized_mut().expect("we just set it")) } - pub(crate) fn initialized_mut( - &mut self, - ) -> Result<&mut UploadQueueInitialized, NotInitialized> { + pub fn initialized_mut(&mut self) -> Result<&mut UploadQueueInitialized, NotInitialized> { use UploadQueue::*; match self { Uninitialized => Err(NotInitialized::Uninitialized), @@ -276,23 +391,23 @@ impl UploadQueue { /// An in-progress upload or delete task. #[derive(Debug)] -pub(crate) struct UploadTask { +pub struct UploadTask { /// Unique ID of this task. Used as the key in `inprogress_tasks` above. - pub(crate) task_id: u64, - pub(crate) retries: AtomicU32, + pub task_id: u64, + pub retries: AtomicU32, - pub(crate) op: UploadOp, + pub op: UploadOp, } /// A deletion of some layers within the lifetime of a timeline. This is not used /// for timeline deletion, which skips this queue and goes directly to DeletionQueue. #[derive(Debug, Clone)] -pub(crate) struct Delete { - pub(crate) layers: Vec<(LayerName, LayerFileMetadata)>, +pub struct Delete { + pub layers: Vec<(LayerName, LayerFileMetadata)>, } -#[derive(Debug)] -pub(crate) enum UploadOp { +#[derive(Clone, Debug)] +pub enum UploadOp { /// Upload a layer file. The last field indicates the last operation for thie file. UploadLayer(ResidentLayer, LayerFileMetadata, Option), @@ -338,3 +453,765 @@ impl std::fmt::Display for UploadOp { } } } + +impl UploadOp { + /// Returns true if self can bypass other, i.e. if the operations don't conflict. index is the + /// active index when other would be uploaded -- if we allow self to bypass other, this would + /// be the active index when self is uploaded. + pub fn can_bypass(&self, other: &UploadOp, index: &IndexPart) -> bool { + match (self, other) { + // Nothing can bypass a barrier or shutdown, and it can't bypass anything. + (UploadOp::Barrier(_), _) | (_, UploadOp::Barrier(_)) => false, + (UploadOp::Shutdown, _) | (_, UploadOp::Shutdown) => false, + + // Uploads and deletes can bypass each other unless they're for the same file. + (UploadOp::UploadLayer(a, ameta, _), UploadOp::UploadLayer(b, bmeta, _)) => { + let aname = &a.layer_desc().layer_name(); + let bname = &b.layer_desc().layer_name(); + !is_same_remote_layer_path(aname, ameta, bname, bmeta) + } + (UploadOp::UploadLayer(u, umeta, _), UploadOp::Delete(d)) + | (UploadOp::Delete(d), UploadOp::UploadLayer(u, umeta, _)) => { + d.layers.iter().all(|(dname, dmeta)| { + !is_same_remote_layer_path(&u.layer_desc().layer_name(), umeta, dname, dmeta) + }) + } + + // Deletes are idempotent and can always bypass each other. + (UploadOp::Delete(_), UploadOp::Delete(_)) => true, + + // Uploads and deletes can bypass an index upload as long as neither the uploaded index + // nor the active index below it references the file. A layer can't be modified or + // deleted while referenced by an index. + // + // Similarly, index uploads can bypass uploads and deletes as long as neither the + // uploaded index nor the active index references the file (the latter would be + // incorrect use by the caller). + (UploadOp::UploadLayer(u, umeta, _), UploadOp::UploadMetadata { uploaded: i }) + | (UploadOp::UploadMetadata { uploaded: i }, UploadOp::UploadLayer(u, umeta, _)) => { + let uname = u.layer_desc().layer_name(); + !i.references(&uname, umeta) && !index.references(&uname, umeta) + } + (UploadOp::Delete(d), UploadOp::UploadMetadata { uploaded: i }) + | (UploadOp::UploadMetadata { uploaded: i }, UploadOp::Delete(d)) => { + d.layers.iter().all(|(dname, dmeta)| { + !i.references(dname, dmeta) && !index.references(dname, dmeta) + }) + } + + // Indexes can never bypass each other. + // TODO: we could coalesce them though, by only uploading the newest ready index. This + // is left for later, out of caution. + (UploadOp::UploadMetadata { .. }, UploadOp::UploadMetadata { .. }) => false, + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + use crate::tenant::harness::{TenantHarness, TIMELINE_ID}; + use crate::tenant::storage_layer::layer::local_layer_path; + use crate::tenant::storage_layer::Layer; + use crate::tenant::Timeline; + use crate::DEFAULT_PG_VERSION; + use itertools::Itertools as _; + use std::str::FromStr as _; + use utils::shard::{ShardCount, ShardIndex, ShardNumber}; + + /// Test helper which asserts that two operations are the same, in lieu of UploadOp PartialEq. + #[track_caller] + fn assert_same_op(a: &UploadOp, b: &UploadOp) { + use UploadOp::*; + match (a, b) { + (UploadLayer(a, ameta, atype), UploadLayer(b, bmeta, btype)) => { + assert_eq!(a.layer_desc().layer_name(), b.layer_desc().layer_name()); + assert_eq!(ameta, bmeta); + assert_eq!(atype, btype); + } + (Delete(a), Delete(b)) => assert_eq!(a.layers, b.layers), + (UploadMetadata { uploaded: a }, UploadMetadata { uploaded: b }) => assert_eq!(a, b), + (Barrier(_), Barrier(_)) => {} + (Shutdown, Shutdown) => {} + (a, b) => panic!("{a:?} != {b:?}"), + } + } + + /// Test helper which asserts that two sets of operations are the same. + #[track_caller] + fn assert_same_ops<'a>( + a: impl IntoIterator, + b: impl IntoIterator, + ) { + a.into_iter() + .zip_eq(b) + .for_each(|(a, b)| assert_same_op(a, b)) + } + + /// Test helper to construct a test timeline. + /// + /// TODO: it really shouldn't be necessary to construct an entire tenant and timeline just to + /// test the upload queue -- decouple ResidentLayer from Timeline. + /// + /// TODO: the upload queue uses TimelineMetadata::example() instead, because there's no way to + /// obtain a TimelineMetadata from a Timeline. + fn make_timeline() -> Arc { + // Grab the current test name from the current thread name. + // TODO: TenantHarness shouldn't take a &'static str, but just leak the test name for now. + let test_name = std::thread::current().name().unwrap().to_string(); + let test_name = Box::leak(test_name.into_boxed_str()); + + let runtime = tokio::runtime::Builder::new_current_thread() + .enable_all() + .build() + .expect("failed to create runtime"); + + runtime + .block_on(async { + let harness = TenantHarness::create(test_name).await?; + let (tenant, ctx) = harness.load().await; + tenant + .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx) + .await + }) + .expect("failed to create timeline") + } + + /// Test helper to construct an (empty) resident layer. + fn make_layer(timeline: &Arc, name: &str) -> ResidentLayer { + make_layer_with_size(timeline, name, 0) + } + + /// Test helper to construct a resident layer with the given size. + fn make_layer_with_size(timeline: &Arc, name: &str, size: usize) -> ResidentLayer { + let metadata = LayerFileMetadata { + generation: timeline.generation, + shard: timeline.get_shard_index(), + file_size: size as u64, + }; + make_layer_with_metadata(timeline, name, metadata) + } + + /// Test helper to construct a layer with the given metadata. + fn make_layer_with_metadata( + timeline: &Arc, + name: &str, + metadata: LayerFileMetadata, + ) -> ResidentLayer { + let name = LayerName::from_str(name).expect("invalid name"); + let local_path = local_layer_path( + timeline.conf, + &timeline.tenant_shard_id, + &timeline.timeline_id, + &name, + &metadata.generation, + ); + std::fs::write(&local_path, vec![0; metadata.file_size as usize]) + .expect("failed to write file"); + Layer::for_resident(timeline.conf, timeline, local_path, name, metadata) + } + + /// Test helper to add a layer to an index and return a new index. + fn index_with(index: &IndexPart, layer: &ResidentLayer) -> Box { + let mut index = index.clone(); + index + .layer_metadata + .insert(layer.layer_desc().layer_name(), layer.metadata()); + Box::new(index) + } + + /// Test helper to remove a layer from an index and return a new index. + fn index_without(index: &IndexPart, layer: &ResidentLayer) -> Box { + let mut index = index.clone(); + index + .layer_metadata + .remove(&layer.layer_desc().layer_name()); + Box::new(index) + } + + /// Nothing can bypass a barrier, and it can't bypass inprogress tasks. + #[test] + fn schedule_barrier() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let tli = make_timeline(); + + let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let (barrier, _) = tokio::sync::watch::channel(()); + + // Enqueue non-conflicting upload, delete, and index before and after a barrier. + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())], + }), + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + UploadOp::Barrier(barrier), + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())], + }), + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + ]; + + queue.queued_operations.extend(ops.clone()); + + // Schedule the initial operations ahead of the barrier. + let tasks = queue.schedule_ready(); + + assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..3]); + assert!(matches!( + queue.queued_operations.front(), + Some(&UploadOp::Barrier(_)) + )); + + // Complete the initial operations. The barrier isn't scheduled while they're pending. + for task in tasks { + assert!(queue.schedule_ready().is_empty()); + queue.complete(task.task_id); + } + + // Schedule the barrier. The later tasks won't schedule until it completes. + let tasks = queue.schedule_ready(); + + assert_eq!(tasks.len(), 1); + assert!(matches!(tasks[0].op, UploadOp::Barrier(_))); + assert_eq!(queue.queued_operations.len(), 3); + + // Complete the barrier. The rest of the tasks schedule immediately. + queue.complete(tasks[0].task_id); + + let tasks = queue.schedule_ready(); + assert_same_ops(tasks.iter().map(|t| &t.op), &ops[4..]); + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Deletes can be scheduled in parallel, even if they're for the same file. + #[test] + fn schedule_delete_parallel() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let tli = make_timeline(); + + // Enqueue a bunch of deletes, some with conflicting names. + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + let ops = [ + UploadOp::Delete(Delete { + layers: vec![(layer0.layer_desc().layer_name(), layer0.metadata())], + }), + UploadOp::Delete(Delete { + layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())], + }), + UploadOp::Delete(Delete { + layers: vec![ + (layer1.layer_desc().layer_name(), layer1.metadata()), + (layer2.layer_desc().layer_name(), layer2.metadata()), + ], + }), + UploadOp::Delete(Delete { + layers: vec![(layer2.layer_desc().layer_name(), layer2.metadata())], + }), + UploadOp::Delete(Delete { + layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())], + }), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Schedule all ready operations. Since deletes don't conflict, they're all scheduled. + let tasks = queue.schedule_ready(); + + assert_same_ops(tasks.iter().map(|t| &t.op), &ops); + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Conflicting uploads are serialized. + #[test] + fn schedule_upload_conflicts() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Enqueue three versions of the same layer, with different file sizes. + let layer0a = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 1); + let layer0b = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 2); + let layer0c = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 3); + + let ops = [ + UploadOp::UploadLayer(layer0a.clone(), layer0a.metadata(), None), + UploadOp::UploadLayer(layer0b.clone(), layer0b.metadata(), None), + UploadOp::UploadLayer(layer0c.clone(), layer0c.metadata(), None), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Only one version should be scheduled and uploaded at a time. + for op in ops { + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &op); + queue.complete(tasks[0].task_id); + } + assert!(queue.schedule_ready().is_empty()); + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Conflicting uploads and deletes are serialized. + #[test] + fn schedule_upload_delete_conflicts() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Enqueue two layer uploads, with a delete of both layers in between them. These should be + // scheduled one at a time, since deletes can't bypass uploads and vice versa. + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![ + (layer0.layer_desc().layer_name(), layer0.metadata()), + (layer1.layer_desc().layer_name(), layer1.metadata()), + ], + }), + UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Only one version should be scheduled and uploaded at a time. + for op in ops { + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &op); + queue.complete(tasks[0].task_id); + } + assert!(queue.schedule_ready().is_empty()); + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Non-conflicting uploads and deletes can bypass the queue, avoiding the conflicting + /// delete/upload operations at the head of the queue. + #[test] + fn schedule_upload_delete_conflicts_bypass() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Enqueue two layer uploads, with a delete of both layers in between them. These should be + // scheduled one at a time, since deletes can't bypass uploads and vice versa. + // + // Also enqueue non-conflicting uploads and deletes at the end. These can bypass the queue + // and run immediately. + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![ + (layer0.layer_desc().layer_name(), layer0.metadata()), + (layer1.layer_desc().layer_name(), layer1.metadata()), + ], + }), + UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None), + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())], + }), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Operations 0, 3, and 4 are scheduled immediately. + let tasks = queue.schedule_ready(); + assert_same_ops(tasks.iter().map(|t| &t.op), [&ops[0], &ops[3], &ops[4]]); + assert_eq!(queue.queued_operations.len(), 2); + + Ok(()) + } + + /// Non-conflicting uploads are parallelized. + #[test] + fn schedule_upload_parallel() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Enqueue three different layer uploads. + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None), + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + ]; + + queue.queued_operations.extend(ops.clone()); + + // All uploads should be scheduled concurrently. + let tasks = queue.schedule_ready(); + + assert_same_ops(tasks.iter().map(|t| &t.op), &ops); + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Index uploads are serialized. + #[test] + fn schedule_index_serial() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + + // Enqueue three uploads of the current empty index. + let index = Box::new(queue.clean.0.clone()); + + let ops = [ + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + ]; + + queue.queued_operations.extend(ops.clone()); + + // The uploads should run serially. + for op in ops { + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &op); + queue.complete(tasks[0].task_id); + } + + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Chains of upload/index operations lead to parallel layer uploads and serial index uploads. + /// This is the common case with layer flushes. + #[test] + fn schedule_index_upload_chain() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Enqueue three uploads of the current empty index. + let index = Box::new(queue.clean.0.clone()); + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let index0 = index_with(&index, &layer0); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let index1 = index_with(&index0, &layer1); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let index2 = index_with(&index1, &layer2); + + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index0.clone(), + }, + UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index1.clone(), + }, + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index2.clone(), + }, + ]; + + queue.queued_operations.extend(ops.clone()); + + // The layer uploads should be scheduled immediately. The indexes must wait. + let upload_tasks = queue.schedule_ready(); + assert_same_ops( + upload_tasks.iter().map(|t| &t.op), + [&ops[0], &ops[2], &ops[4]], + ); + + // layer2 completes first. None of the indexes can upload yet. + queue.complete(upload_tasks[2].task_id); + assert!(queue.schedule_ready().is_empty()); + + // layer0 completes. index0 can upload. It completes. + queue.complete(upload_tasks[0].task_id); + let index_tasks = queue.schedule_ready(); + assert_eq!(index_tasks.len(), 1); + assert_same_op(&index_tasks[0].op, &ops[1]); + queue.complete(index_tasks[0].task_id); + + // layer 1 completes. This unblocks index 1 then index 2. + queue.complete(upload_tasks[1].task_id); + + let index_tasks = queue.schedule_ready(); + assert_eq!(index_tasks.len(), 1); + assert_same_op(&index_tasks[0].op, &ops[3]); + queue.complete(index_tasks[0].task_id); + + let index_tasks = queue.schedule_ready(); + assert_eq!(index_tasks.len(), 1); + assert_same_op(&index_tasks[0].op, &ops[5]); + queue.complete(index_tasks[0].task_id); + + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// A delete can't bypass an index upload if an index ahead of it still references it. + #[test] + fn schedule_index_delete_dereferenced() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Create a layer to upload. + let layer = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let index_upload = index_with(&queue.clean.0, &layer); + + // Remove the layer reference in a new index, then delete the layer. + let index_deref = index_without(&index_upload, &layer); + + let ops = [ + // Initial upload. + UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index_upload.clone(), + }, + // Dereference the layer and delete it. + UploadOp::UploadMetadata { + uploaded: index_deref.clone(), + }, + UploadOp::Delete(Delete { + layers: vec![(layer.layer_desc().layer_name(), layer.metadata())], + }), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Operations are serialized. + for op in ops { + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &op); + queue.complete(tasks[0].task_id); + } + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// An upload with a reused layer name doesn't clobber the previous layer. Specifically, a + /// dereference/upload/reference cycle can't allow the upload to bypass the reference. + #[test] + fn schedule_index_upload_dereferenced() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let tli = make_timeline(); + + // Create a layer to upload. + let layer = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + // Upload the layer. Then dereference the layer, and upload/reference it again. + let index_upload = index_with(&queue.clean.0, &layer); + let index_deref = index_without(&index_upload, &layer); + let index_ref = index_with(&index_deref, &layer); + + let ops = [ + // Initial upload. + UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index_upload.clone(), + }, + // Dereference the layer. + UploadOp::UploadMetadata { + uploaded: index_deref.clone(), + }, + // Replace and reference the layer. + UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), + UploadOp::UploadMetadata { + uploaded: index_ref.clone(), + }, + ]; + + queue.queued_operations.extend(ops.clone()); + + // Operations are serialized. + for op in ops { + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &op); + queue.complete(tasks[0].task_id); + } + assert!(queue.queued_operations.is_empty()); + + Ok(()) + } + + /// Nothing can bypass a shutdown, and it waits for inprogress tasks. It's never returned from + /// next_ready(), but is left at the head of the queue. + #[test] + fn schedule_shutdown() -> anyhow::Result<()> { + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let tli = make_timeline(); + + let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + // Enqueue non-conflicting upload, delete, and index before and after a shutdown. + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())], + }), + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + UploadOp::Shutdown, + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())], + }), + UploadOp::UploadMetadata { + uploaded: index.clone(), + }, + ]; + + queue.queued_operations.extend(ops.clone()); + + // Schedule the initial operations ahead of the shutdown. + let tasks = queue.schedule_ready(); + + assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..3]); + assert!(matches!( + queue.queued_operations.front(), + Some(&UploadOp::Shutdown) + )); + + // Complete the initial operations. The shutdown isn't triggered while they're pending. + for task in tasks { + assert!(queue.schedule_ready().is_empty()); + queue.complete(task.task_id); + } + + // The shutdown is triggered the next time we try to pull an operation. It isn't returned, + // but is left in the queue. + assert!(!queue.shutdown_ready.is_closed()); + assert!(queue.next_ready().is_none()); + assert!(queue.shutdown_ready.is_closed()); + + Ok(()) + } + + /// Tests that can_bypass takes name, generation and shard index into account for all operations. + #[test] + fn can_bypass_path() -> anyhow::Result<()> { + let tli = make_timeline(); + + let name0 = &"000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"; + let name1 = &"100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"; + + // Asserts that layers a and b either can or can't bypass each other, for all combinations + // of operations (except Delete and UploadMetadata which are special-cased). + #[track_caller] + fn assert_can_bypass(a: ResidentLayer, b: ResidentLayer, can_bypass: bool) { + let index = IndexPart::empty(TimelineMetadata::example()); + for (a, b) in make_ops(a).into_iter().zip(make_ops(b)) { + match (&a, &b) { + // Deletes can always bypass each other. + (UploadOp::Delete(_), UploadOp::Delete(_)) => assert!(a.can_bypass(&b, &index)), + // Indexes can never bypass each other. + (UploadOp::UploadMetadata { .. }, UploadOp::UploadMetadata { .. }) => { + assert!(!a.can_bypass(&b, &index)) + } + // For other operations, assert as requested. + (a, b) => assert_eq!(a.can_bypass(b, &index), can_bypass), + } + } + } + + fn make_ops(layer: ResidentLayer) -> Vec { + let mut index = IndexPart::empty(TimelineMetadata::example()); + index + .layer_metadata + .insert(layer.layer_desc().layer_name(), layer.metadata()); + vec![ + UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), + UploadOp::Delete(Delete { + layers: vec![(layer.layer_desc().layer_name(), layer.metadata())], + }), + UploadOp::UploadMetadata { + uploaded: Box::new(index), + }, + ] + } + + // Makes a ResidentLayer. + let layer = |name: &'static str, shard: Option, generation: u32| -> ResidentLayer { + let shard = shard + .map(|n| ShardIndex::new(ShardNumber(n), ShardCount(8))) + .unwrap_or(ShardIndex::unsharded()); + let metadata = LayerFileMetadata { + shard, + generation: Generation::Valid(generation), + file_size: 0, + }; + make_layer_with_metadata(&tli, name, metadata) + }; + + // Same name and metadata can't bypass. This goes both for unsharded and sharded, as well as + // 0 or >0 generation. + assert_can_bypass(layer(name0, None, 0), layer(name0, None, 0), false); + assert_can_bypass(layer(name0, Some(0), 0), layer(name0, Some(0), 0), false); + assert_can_bypass(layer(name0, None, 1), layer(name0, None, 1), false); + + // Different names can bypass. + assert_can_bypass(layer(name0, None, 0), layer(name1, None, 0), true); + + // Different shards can bypass. Shard 0 is different from unsharded. + assert_can_bypass(layer(name0, Some(0), 0), layer(name0, Some(1), 0), true); + assert_can_bypass(layer(name0, Some(0), 0), layer(name0, None, 0), true); + + // Different generations can bypass, both sharded and unsharded. + assert_can_bypass(layer(name0, None, 0), layer(name0, None, 1), true); + assert_can_bypass(layer(name0, Some(1), 0), layer(name0, Some(1), 1), true); + + Ok(()) + } +} From d36112d20fdc249c324308f82b2b81bb124065d9 Mon Sep 17 00:00:00 2001 From: Heikki Linnakangas Date: Tue, 14 Jan 2025 19:02:35 +0200 Subject: [PATCH 06/31] Simplify compute dockerfile by setting PATH just once (#10357) By setting PATH in the 'pg-build' layer, all the extension build layers will inherit. No need to pass PG_CONFIG to all the various make invocations either: once pg_config is in PATH, the Makefiles will pick it up from there. --- compute/compute-node.Dockerfile | 168 ++++++++++++-------------------- 1 file changed, 63 insertions(+), 105 deletions(-) diff --git a/compute/compute-node.Dockerfile b/compute/compute-node.Dockerfile index 89cee6761f..299f4444a3 100644 --- a/compute/compute-node.Dockerfile +++ b/compute/compute-node.Dockerfile @@ -104,16 +104,18 @@ RUN cd postgres && \ esac; \ done; +# Set PATH for all the subsequent build steps +ENV PATH="/usr/local/pgsql/bin:$PATH" + ######################################################################################### # # Layer "postgis-build" # Build PostGIS from the upstream PostGIS mirror. # ######################################################################################### -FROM build-deps AS postgis-build +FROM pg-build AS postgis-build ARG DEBIAN_VERSION ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ RUN apt update && \ apt install --no-install-recommends --no-install-suggests -y \ gdal-bin libboost-dev libboost-thread-dev libboost-filesystem-dev \ @@ -151,8 +153,6 @@ RUN case "${DEBIAN_VERSION}" in \ DESTDIR=/sfcgal ninja install -j $(getconf _NPROCESSORS_ONLN) && \ ninja clean && cp -R /sfcgal/* / -ENV PATH="/usr/local/pgsql/bin:$PATH" - # Postgis 3.5.0 supports v17 RUN case "${PG_VERSION}" in \ "v17") \ @@ -227,9 +227,8 @@ RUN case "${PG_VERSION}" in \ # Build plv8 # ######################################################################################### -FROM build-deps AS plv8-build +FROM pg-build AS plv8-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ COPY compute/patches/plv8-3.1.10.patch /plv8-3.1.10.patch @@ -264,7 +263,6 @@ RUN case "${PG_VERSION}" in \ # generate and copy upgrade scripts mkdir -p upgrade && ./generate_upgrade.sh ${PLV8_TAG#v} && \ cp upgrade/* /usr/local/pgsql/share/extension/ && \ - export PATH="/usr/local/pgsql/bin:$PATH" && \ make DOCKER=1 -j $(getconf _NPROCESSORS_ONLN) install && \ rm -rf /plv8-* && \ find /usr/local/pgsql/ -name "plv8-*.so" | xargs strip && \ @@ -291,9 +289,8 @@ RUN case "${PG_VERSION}" in \ # Build h3_pg # ######################################################################################### -FROM build-deps AS h3-pg-build +FROM pg-build AS h3-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v4.1.0 - Jan 18, 2023 @@ -314,7 +311,6 @@ RUN mkdir -p /h3/usr/ && \ RUN wget https://github.com/zachasme/h3-pg/archive/refs/tags/v4.1.3.tar.gz -O h3-pg.tar.gz && \ echo "5c17f09a820859ffe949f847bebf1be98511fb8f1bd86f94932512c00479e324 h3-pg.tar.gz" | sha256sum --check && \ mkdir h3-pg-src && cd h3-pg-src && tar xzf ../h3-pg.tar.gz --strip-components=1 -C . && \ - export PATH="/usr/local/pgsql/bin:$PATH" && \ make -j $(getconf _NPROCESSORS_ONLN) && \ make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/h3.control && \ @@ -326,17 +322,16 @@ RUN wget https://github.com/zachasme/h3-pg/archive/refs/tags/v4.1.3.tar.gz -O h3 # compile unit extension # ######################################################################################### -FROM build-deps AS unit-pg-build +FROM pg-build AS unit-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release 7.9 - Sep 15, 2024 RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.9.tar.gz -O postgresql-unit.tar.gz && \ echo "e46de6245dcc8b2c2ecf29873dbd43b2b346773f31dd5ce4b8315895a052b456 postgresql-unit.tar.gz" | sha256sum --check && \ mkdir postgresql-unit-src && cd postgresql-unit-src && tar xzf ../postgresql-unit.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ # unit extension's "create extension" script relies on absolute install path to fill some reference tables. # We move the extension from '/usr/local/pgsql/' to '/usr/local/' after it is build. So we need to adjust the path. # This one-liner removes pgsql/ part of the path. @@ -350,9 +345,8 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.9.tar.gz - # compile pgvector extension # ######################################################################################### -FROM build-deps AS vector-pg-build +FROM pg-build AS vector-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ COPY compute/patches/pgvector.patch /pgvector.patch @@ -366,8 +360,8 @@ RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.8.0.tar.gz -O echo "867a2c328d4928a5a9d6f052cd3bc78c7d60228a9b914ad32aa3db88e9de27b0 pgvector.tar.gz" | sha256sum --check && \ mkdir pgvector-src && cd pgvector-src && tar xzf ../pgvector.tar.gz --strip-components=1 -C . && \ patch -p1 < /pgvector.patch && \ - make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" && \ + make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/vector.control ######################################################################################### @@ -376,16 +370,15 @@ RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.8.0.tar.gz -O # compile pgjwt extension # ######################################################################################### -FROM build-deps AS pgjwt-pg-build +FROM pg-build AS pgjwt-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # doesn't use releases, last commit f3d82fd - Mar 2, 2023 RUN wget https://github.com/michelp/pgjwt/archive/f3d82fd30151e754e19ce5d6a06c71c20689ce3d.tar.gz -O pgjwt.tar.gz && \ echo "dae8ed99eebb7593b43013f6532d772b12dfecd55548d2673f2dfd0163f6d2b9 pgjwt.tar.gz" | sha256sum --check && \ mkdir pgjwt-src && cd pgjwt-src && tar xzf ../pgjwt.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgjwt.control ######################################################################################### @@ -394,17 +387,16 @@ RUN wget https://github.com/michelp/pgjwt/archive/f3d82fd30151e754e19ce5d6a06c71 # compile hypopg extension # ######################################################################################### -FROM build-deps AS hypopg-pg-build +FROM pg-build AS hypopg-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # HypoPG 1.4.1 supports v17 # last release 1.4.1 - Apr 28, 2024 RUN wget https://github.com/HypoPG/hypopg/archive/refs/tags/1.4.1.tar.gz -O hypopg.tar.gz && \ echo "9afe6357fd389d8d33fad81703038ce520b09275ec00153c6c89282bcdedd6bc hypopg.tar.gz" | sha256sum --check && \ mkdir hypopg-src && cd hypopg-src && tar xzf ../hypopg.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/hypopg.control ######################################################################################### @@ -413,17 +405,16 @@ RUN wget https://github.com/HypoPG/hypopg/archive/refs/tags/1.4.1.tar.gz -O hypo # compile pg_hashids extension # ######################################################################################### -FROM build-deps AS pg-hashids-pg-build +FROM pg-build AS pg-hashids-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v1.2.1 -Jan 12, 2018 RUN wget https://github.com/iCyberon/pg_hashids/archive/refs/tags/v1.2.1.tar.gz -O pg_hashids.tar.gz && \ echo "74576b992d9277c92196dd8d816baa2cc2d8046fe102f3dcd7f3c3febed6822a pg_hashids.tar.gz" | sha256sum --check && \ mkdir pg_hashids-src && cd pg_hashids-src && tar xzf ../pg_hashids.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) install USE_PGXS=1 && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_hashids.control ######################################################################################### @@ -432,9 +423,8 @@ RUN wget https://github.com/iCyberon/pg_hashids/archive/refs/tags/v1.2.1.tar.gz # compile rum extension # ######################################################################################### -FROM build-deps AS rum-pg-build +FROM pg-build AS rum-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ COPY compute/patches/rum.patch /rum.patch @@ -445,8 +435,8 @@ RUN wget https://github.com/postgrespro/rum/archive/cb1edffc57736cd2a4455f8d0fea echo "65e0a752e99f4c3226400c9b899f997049e93503db8bf5c8072efa136d32fd83 rum.tar.gz" | sha256sum --check && \ mkdir rum-src && cd rum-src && tar xzf ../rum.tar.gz --strip-components=1 -C . && \ patch -p1 < /rum.patch && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) install USE_PGXS=1 && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/rum.control ######################################################################################### @@ -455,17 +445,16 @@ RUN wget https://github.com/postgrespro/rum/archive/cb1edffc57736cd2a4455f8d0fea # compile pgTAP extension # ######################################################################################### -FROM build-deps AS pgtap-pg-build +FROM pg-build AS pgtap-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # pgtap 1.3.3 supports v17 # last release v1.3.3 - Apr 8, 2024 RUN wget https://github.com/theory/pgtap/archive/refs/tags/v1.3.3.tar.gz -O pgtap.tar.gz && \ echo "325ea79d0d2515bce96bce43f6823dcd3effbd6c54cb2a4d6c2384fffa3a14c7 pgtap.tar.gz" | sha256sum --check && \ mkdir pgtap-src && cd pgtap-src && tar xzf ../pgtap.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgtap.control ######################################################################################### @@ -474,17 +463,16 @@ RUN wget https://github.com/theory/pgtap/archive/refs/tags/v1.3.3.tar.gz -O pgta # compile ip4r extension # ######################################################################################### -FROM build-deps AS ip4r-pg-build +FROM pg-build AS ip4r-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v2.4.2 - Jul 29, 2023 RUN wget https://github.com/RhodiumToad/ip4r/archive/refs/tags/2.4.2.tar.gz -O ip4r.tar.gz && \ echo "0f7b1f159974f49a47842a8ab6751aecca1ed1142b6d5e38d81b064b2ead1b4b ip4r.tar.gz" | sha256sum --check && \ mkdir ip4r-src && cd ip4r-src && tar xzf ../ip4r.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/ip4r.control ######################################################################################### @@ -493,17 +481,16 @@ RUN wget https://github.com/RhodiumToad/ip4r/archive/refs/tags/2.4.2.tar.gz -O i # compile Prefix extension # ######################################################################################### -FROM build-deps AS prefix-pg-build +FROM pg-build AS prefix-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v1.2.10 - Jul 5, 2023 RUN wget https://github.com/dimitri/prefix/archive/refs/tags/v1.2.10.tar.gz -O prefix.tar.gz && \ echo "4342f251432a5f6fb05b8597139d3ccde8dcf87e8ca1498e7ee931ca057a8575 prefix.tar.gz" | sha256sum --check && \ mkdir prefix-src && cd prefix-src && tar xzf ../prefix.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/prefix.control ######################################################################################### @@ -512,17 +499,16 @@ RUN wget https://github.com/dimitri/prefix/archive/refs/tags/v1.2.10.tar.gz -O p # compile hll extension # ######################################################################################### -FROM build-deps AS hll-pg-build +FROM pg-build AS hll-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v2.18 - Aug 29, 2023 RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.18.tar.gz -O hll.tar.gz && \ echo "e2f55a6f4c4ab95ee4f1b4a2b73280258c5136b161fe9d059559556079694f0e hll.tar.gz" | sha256sum --check && \ mkdir hll-src && cd hll-src && tar xzf ../hll.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/hll.control ######################################################################################### @@ -531,17 +517,16 @@ RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.18.tar # compile plpgsql_check extension # ######################################################################################### -FROM build-deps AS plpgsql-check-pg-build +FROM pg-build AS plpgsql-check-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # plpgsql_check v2.7.11 supports v17 # last release v2.7.11 - Sep 16, 2024 RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.7.11.tar.gz -O plpgsql_check.tar.gz && \ echo "208933f8dbe8e0d2628eb3851e9f52e6892b8e280c63700c0f1ce7883625d172 plpgsql_check.tar.gz" | sha256sum --check && \ mkdir plpgsql_check-src && cd plpgsql_check-src && tar xzf ../plpgsql_check.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) USE_PGXS=1 && \ + make -j $(getconf _NPROCESSORS_ONLN) install USE_PGXS=1 && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/plpgsql_check.control ######################################################################################### @@ -550,11 +535,8 @@ RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.7.11.tar.gz # compile timescaledb extension # ######################################################################################### -FROM build-deps AS timescaledb-pg-build -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ - +FROM pg-build AS timescaledb-pg-build ARG PG_VERSION -ENV PATH="/usr/local/pgsql/bin:$PATH" RUN case "${PG_VERSION}" in \ "v14" | "v15") \ @@ -585,11 +567,8 @@ RUN case "${PG_VERSION}" in \ # compile pg_hint_plan extension # ######################################################################################### -FROM build-deps AS pg-hint-plan-pg-build -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ - +FROM pg-build AS pg-hint-plan-pg-build ARG PG_VERSION -ENV PATH="/usr/local/pgsql/bin:$PATH" # version-specific, has separate releases for each version RUN case "${PG_VERSION}" in \ @@ -627,14 +606,12 @@ RUN case "${PG_VERSION}" in \ # compile pg_cron extension # ######################################################################################### -FROM build-deps AS pg-cron-pg-build +FROM pg-build AS pg-cron-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # This is an experimental extension that we do not support on prod yet. # !Do not remove! # We set it in shared_preload_libraries and computes will fail to start if library is not found. -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.6.4.tar.gz -O pg_cron.tar.gz && \ echo "52d1850ee7beb85a4cb7185731ef4e5a90d1de216709d8988324b0d02e76af61 pg_cron.tar.gz" | sha256sum --check && \ mkdir pg_cron-src && cd pg_cron-src && tar xzf ../pg_cron.tar.gz --strip-components=1 -C . && \ @@ -648,9 +625,8 @@ RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.6.4.tar.gz -O # compile rdkit extension # ######################################################################################### -FROM build-deps AS rdkit-pg-build +FROM pg-build AS rdkit-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ RUN apt update && \ apt install --no-install-recommends --no-install-suggests -y \ @@ -668,7 +644,13 @@ RUN apt update && \ # Use new version only for v17 # because Release_2024_09_1 has some backward incompatible changes # https://github.com/rdkit/rdkit/releases/tag/Release_2024_09_1 -ENV PATH="/usr/local/pgsql/bin/:/usr/local/pgsql/:$PATH" + +# XXX: /usr/local/pgsql/bin is already in PATH, and that should be enough to find +# pg_config. For some reason the rdkit cmake script doesn't work with just that, +# however. By also adding /usr/local/pgsql, it works, which is weird because there +# are no executables in that directory. +ENV PATH="/usr/local/pgsql:$PATH" + RUN case "${PG_VERSION}" in \ "v17") \ export RDKIT_VERSION=Release_2024_09_1 \ @@ -721,13 +703,11 @@ RUN case "${PG_VERSION}" in \ # compile pg_uuidv7 extension # ######################################################################################### -FROM build-deps AS pg-uuidv7-pg-build +FROM pg-build AS pg-uuidv7-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v1.6.0 - Oct 9, 2024 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/fboulnois/pg_uuidv7/archive/refs/tags/v1.6.0.tar.gz -O pg_uuidv7.tar.gz && \ echo "0fa6c710929d003f6ce276a7de7a864e9d1667b2d78be3dc2c07f2409eb55867 pg_uuidv7.tar.gz" | sha256sum --check && \ mkdir pg_uuidv7-src && cd pg_uuidv7-src && tar xzf ../pg_uuidv7.tar.gz --strip-components=1 -C . && \ @@ -741,13 +721,11 @@ RUN wget https://github.com/fboulnois/pg_uuidv7/archive/refs/tags/v1.6.0.tar.gz # compile pg_roaringbitmap extension # ######################################################################################### -FROM build-deps AS pg-roaringbitmap-pg-build +FROM pg-build AS pg-roaringbitmap-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # not version-specific # last release v0.5.4 - Jun 28, 2022 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/ChenHuajun/pg_roaringbitmap/archive/refs/tags/v0.5.4.tar.gz -O pg_roaringbitmap.tar.gz && \ echo "b75201efcb1c2d1b014ec4ae6a22769cc7a224e6e406a587f5784a37b6b5a2aa pg_roaringbitmap.tar.gz" | sha256sum --check && \ mkdir pg_roaringbitmap-src && cd pg_roaringbitmap-src && tar xzf ../pg_roaringbitmap.tar.gz --strip-components=1 -C . && \ @@ -761,16 +739,14 @@ RUN wget https://github.com/ChenHuajun/pg_roaringbitmap/archive/refs/tags/v0.5.4 # compile pg_semver extension # ######################################################################################### -FROM build-deps AS pg-semver-pg-build +FROM pg-build AS pg-semver-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # Release 0.40.0 breaks backward compatibility with previous versions # see release note https://github.com/theory/pg-semver/releases/tag/v0.40.0 # Use new version only for v17 # # last release v0.40.0 - Jul 22, 2024 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN case "${PG_VERSION}" in \ "v17") \ export SEMVER_VERSION=0.40.0 \ @@ -797,13 +773,11 @@ RUN case "${PG_VERSION}" in \ # compile pg_embedding extension # ######################################################################################### -FROM build-deps AS pg-embedding-pg-build -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ +FROM pg-build AS pg-embedding-pg-build # This is our extension, support stopped in favor of pgvector # TODO: deprecate it ARG PG_VERSION -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN case "${PG_VERSION}" in \ "v14" | "v15") \ export PG_EMBEDDING_VERSION=0.3.5 \ @@ -824,20 +798,18 @@ RUN case "${PG_VERSION}" in \ # compile anon extension # ######################################################################################### -FROM build-deps AS pg-anon-pg-build +FROM pg-build AS pg-anon-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # This is an experimental extension, never got to real production. # !Do not remove! It can be present in shared_preload_libraries and compute will fail to start if library is not found. -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN case "${PG_VERSION}" in "v17") \ echo "postgresql_anonymizer does not yet support PG17" && exit 0;; \ esac && \ wget https://github.com/neondatabase/postgresql_anonymizer/archive/refs/tags/neon_1.1.1.tar.gz -O pg_anon.tar.gz && \ echo "321ea8d5c1648880aafde850a2c576e4a9e7b9933a34ce272efc839328999fa9 pg_anon.tar.gz" | sha256sum --check && \ mkdir pg_anon-src && cd pg_anon-src && tar xzf ../pg_anon.tar.gz --strip-components=1 -C . && \ - make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \ + make -j $(getconf _NPROCESSORS_ONLN) install && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/anon.control ######################################################################################### @@ -846,9 +818,8 @@ RUN case "${PG_VERSION}" in "v17") \ # This layer is used to build `pgrx` deps # ######################################################################################### -FROM build-deps AS rust-extensions-build +FROM pg-build AS rust-extensions-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ RUN apt update && \ apt install --no-install-recommends --no-install-suggests -y curl libclang-dev && \ @@ -856,7 +827,7 @@ RUN apt update && \ useradd -ms /bin/bash nonroot -b /home ENV HOME=/home/nonroot -ENV PATH="/home/nonroot/.cargo/bin:/usr/local/pgsql/bin/:$PATH" +ENV PATH="/home/nonroot/.cargo/bin:$PATH" USER nonroot WORKDIR /home/nonroot @@ -883,9 +854,8 @@ USER root # and eventually get merged with `rust-extensions-build` # ######################################################################################### -FROM build-deps AS rust-extensions-build-pgrx12 +FROM pg-build AS rust-extensions-build-pgrx12 ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ RUN apt update && \ apt install --no-install-recommends --no-install-suggests -y curl libclang-dev && \ @@ -893,7 +863,7 @@ RUN apt update && \ useradd -ms /bin/bash nonroot -b /home ENV HOME=/home/nonroot -ENV PATH="/home/nonroot/.cargo/bin:/usr/local/pgsql/bin/:$PATH" +ENV PATH="/home/nonroot/.cargo/bin:$PATH" USER nonroot WORKDIR /home/nonroot @@ -1068,13 +1038,11 @@ RUN wget https://github.com/neondatabase/pg_session_jwt/archive/refs/tags/v0.2.0 # ######################################################################################### -FROM build-deps AS wal2json-pg-build +FROM pg-build AS wal2json-pg-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # wal2json wal2json_2_6 supports v17 # last release wal2json_2_6 - Apr 25, 2024 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/eulerto/wal2json/archive/refs/tags/wal2json_2_6.tar.gz -O wal2json.tar.gz && \ echo "18b4bdec28c74a8fc98a11c72de38378a760327ef8e5e42e975b0029eb96ba0d wal2json.tar.gz" | sha256sum --check && \ mkdir wal2json-src && cd wal2json-src && tar xzf ../wal2json.tar.gz --strip-components=1 -C . && \ @@ -1087,13 +1055,11 @@ RUN wget https://github.com/eulerto/wal2json/archive/refs/tags/wal2json_2_6.tar. # compile pg_ivm extension # ######################################################################################### -FROM build-deps AS pg-ivm-build +FROM pg-build AS pg-ivm-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # pg_ivm v1.9 supports v17 # last release v1.9 - Jul 31 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/sraoss/pg_ivm/archive/refs/tags/v1.9.tar.gz -O pg_ivm.tar.gz && \ echo "59e15722939f274650abf637f315dd723c87073496ca77236b044cb205270d8b pg_ivm.tar.gz" | sha256sum --check && \ mkdir pg_ivm-src && cd pg_ivm-src && tar xzf ../pg_ivm.tar.gz --strip-components=1 -C . && \ @@ -1107,13 +1073,11 @@ RUN wget https://github.com/sraoss/pg_ivm/archive/refs/tags/v1.9.tar.gz -O pg_iv # compile pg_partman extension # ######################################################################################### -FROM build-deps AS pg-partman-build +FROM pg-build AS pg-partman-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ # should support v17 https://github.com/pgpartman/pg_partman/discussions/693 # last release 5.1.0 Apr 2, 2024 -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/pgpartman/pg_partman/archive/refs/tags/v5.1.0.tar.gz -O pg_partman.tar.gz && \ echo "3e3a27d7ff827295d5c55ef72f07a49062d6204b3cb0b9a048645d6db9f3cb9f pg_partman.tar.gz" | sha256sum --check && \ mkdir pg_partman-src && cd pg_partman-src && tar xzf ../pg_partman.tar.gz --strip-components=1 -C . && \ @@ -1129,9 +1093,6 @@ RUN wget https://github.com/pgpartman/pg_partman/archive/refs/tags/v5.1.0.tar.gz ######################################################################################### FROM rust-extensions-build AS pg-mooncake-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ - -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/Mooncake-Labs/pg_mooncake/releases/download/v0.1.0/pg_mooncake-0.1.0.tar.gz -O pg_mooncake.tar.gz && \ echo "eafd059b77f541f11525eb8affcd66a176968cbd8fe7c0d436e733f2aa4da59f pg_mooncake.tar.gz" | sha256sum --check && \ @@ -1147,11 +1108,8 @@ RUN wget https://github.com/Mooncake-Labs/pg_mooncake/releases/download/v0.1.0/p # ######################################################################################### -FROM build-deps AS pg-repack-build +FROM pg-build AS pg-repack-build ARG PG_VERSION -COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/ - -ENV PATH="/usr/local/pgsql/bin/:$PATH" RUN wget https://github.com/reorg/pg_repack/archive/refs/tags/ver_1.5.2.tar.gz -O pg_repack.tar.gz && \ echo '4516cad42251ed3ad53ff619733004db47d5755acac83f75924cd94d1c4fb681 pg_repack.tar.gz' | sha256sum --check && \ From e58e29e63994373e5645a5123e0be191bf693813 Mon Sep 17 00:00:00 2001 From: Erik Grinaker Date: Tue, 14 Jan 2025 19:01:14 +0100 Subject: [PATCH 07/31] pageserver: limit number of upload queue tasks (#10384) ## Problem The upload queue can currently schedule an arbitrary number of tasks. This can both spawn an unbounded number of Tokio tasks, and also significantly slow down upload queue scheduling as it's quadratic in number of operations. Touches #10096. ## Summary of changes Limit the number of inprogress tasks to the remote storage upload concurrency. While this concurrency limit is shared across all tenants, there's certainly no point in scheduling more than this -- we could even consider setting the limit lower, but don't for now to avoid artificially constraining tenants. --- libs/remote_storage/src/config.rs | 11 +++ pageserver/benches/upload_queue.rs | 2 +- .../src/tenant/remote_timeline_client.rs | 29 +++++++- pageserver/src/tenant/upload_queue.rs | 73 ++++++++++++++++--- 4 files changed, 99 insertions(+), 16 deletions(-) diff --git a/libs/remote_storage/src/config.rs b/libs/remote_storage/src/config.rs index 49b1d9dc87..dae141bf77 100644 --- a/libs/remote_storage/src/config.rs +++ b/libs/remote_storage/src/config.rs @@ -43,6 +43,17 @@ impl RemoteStorageKind { } } +impl RemoteStorageConfig { + /// Helper to fetch the configured concurrency limit. + pub fn concurrency_limit(&self) -> Option { + match &self.storage { + RemoteStorageKind::LocalFs { .. } => None, + RemoteStorageKind::AwsS3(c) => Some(c.concurrency_limit.into()), + RemoteStorageKind::AzureContainer(c) => Some(c.concurrency_limit.into()), + } + } +} + fn default_timeout() -> Duration { RemoteStorageConfig::DEFAULT_TIMEOUT } diff --git a/pageserver/benches/upload_queue.rs b/pageserver/benches/upload_queue.rs index 528b3d5490..ed644b0e3c 100644 --- a/pageserver/benches/upload_queue.rs +++ b/pageserver/benches/upload_queue.rs @@ -53,7 +53,7 @@ fn bench_upload_queue_next_ready(c: &mut Criterion) { // Construct the queue. let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&index)?; + let queue = queue.initialize_with_current_remote_index_part(&index, 0)?; // Populate inprogress_tasks with a bunch of layer1 deletions. let delete = UploadOp::Delete(Delete { diff --git a/pageserver/src/tenant/remote_timeline_client.rs b/pageserver/src/tenant/remote_timeline_client.rs index 75e8da496d..1602765585 100644 --- a/pageserver/src/tenant/remote_timeline_client.rs +++ b/pageserver/src/tenant/remote_timeline_client.rs @@ -425,8 +425,16 @@ impl RemoteTimelineClient { /// an index file upload, i.e., it's not empty. /// The given `index_part` must be the one on the remote. pub fn init_upload_queue(&self, index_part: &IndexPart) -> anyhow::Result<()> { + // Set the maximum number of inprogress tasks to the remote storage concurrency. There's + // certainly no point in starting more upload tasks than this. + let inprogress_limit = self + .conf + .remote_storage_config + .as_ref() + .and_then(|r| r.concurrency_limit()) + .unwrap_or(0); let mut upload_queue = self.upload_queue.lock().unwrap(); - upload_queue.initialize_with_current_remote_index_part(index_part)?; + upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?; self.update_remote_physical_size_gauge(Some(index_part)); info!( "initialized upload queue from remote index with {} layer files", @@ -441,8 +449,16 @@ impl RemoteTimelineClient { &self, local_metadata: &TimelineMetadata, ) -> anyhow::Result<()> { + // Set the maximum number of inprogress tasks to the remote storage concurrency. There's + // certainly no point in starting more upload tasks than this. + let inprogress_limit = self + .conf + .remote_storage_config + .as_ref() + .and_then(|r| r.concurrency_limit()) + .unwrap_or(0); let mut upload_queue = self.upload_queue.lock().unwrap(); - upload_queue.initialize_empty_remote(local_metadata)?; + upload_queue.initialize_empty_remote(local_metadata, inprogress_limit)?; self.update_remote_physical_size_gauge(None); info!("initialized upload queue as empty"); Ok(()) @@ -458,9 +474,15 @@ impl RemoteTimelineClient { let deleted_at = index_part.deleted_at.ok_or(anyhow::anyhow!( "bug: it is responsibility of the caller to provide index part from MaybeDeletedIndexPart::Deleted" ))?; + let inprogress_limit = self + .conf + .remote_storage_config + .as_ref() + .and_then(|r| r.concurrency_limit()) + .unwrap_or(0); let mut upload_queue = self.upload_queue.lock().unwrap(); - upload_queue.initialize_with_current_remote_index_part(index_part)?; + upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?; self.update_remote_physical_size_gauge(Some(index_part)); self.stop_impl(&mut upload_queue); @@ -2355,6 +2377,7 @@ impl RemoteTimelineClient { // but for this use case it doesnt really makes sense to bring unsafe code only for this usage point. // Deletion is not really perf sensitive so there shouldnt be any problems with cloning a fraction of it. let upload_queue_for_deletion = UploadQueueInitialized { + inprogress_limit: initialized.inprogress_limit, task_counter: 0, dirty: initialized.dirty.clone(), clean: initialized.clean.clone(), diff --git a/pageserver/src/tenant/upload_queue.rs b/pageserver/src/tenant/upload_queue.rs index bd524e8153..09c8f6ad8c 100644 --- a/pageserver/src/tenant/upload_queue.rs +++ b/pageserver/src/tenant/upload_queue.rs @@ -51,6 +51,9 @@ pub enum OpType { /// This keeps track of queued and in-progress tasks. pub struct UploadQueueInitialized { + /// Maximum number of inprogress tasks to schedule. 0 is no limit. + pub(crate) inprogress_limit: usize, + /// Counter to assign task IDs pub(crate) task_counter: u64, @@ -128,8 +131,14 @@ impl UploadQueueInitialized { /// the queue if it doesn't conflict with operations ahead of it. /// /// None may be returned even if the queue isn't empty, if no operations are ready yet. + /// + /// NB: this is quadratic, but queues are expected to be small, and bounded by inprogress_limit. pub fn next_ready(&mut self) -> Option { - // NB: this is quadratic, but queues are expected to be small. + // If inprogress_tasks is already at limit, don't schedule anything more. + if self.inprogress_limit > 0 && self.inprogress_tasks.len() >= self.inprogress_limit { + return None; + } + for (i, candidate) in self.queued_operations.iter().enumerate() { // If this candidate is ready, go for it. Otherwise, try the next one. if self.is_ready(i) { @@ -289,6 +298,7 @@ impl UploadQueue { pub fn initialize_empty_remote( &mut self, metadata: &TimelineMetadata, + inprogress_limit: usize, ) -> anyhow::Result<&mut UploadQueueInitialized> { match self { UploadQueue::Uninitialized => (), @@ -302,6 +312,7 @@ impl UploadQueue { let index_part = IndexPart::empty(metadata.clone()); let state = UploadQueueInitialized { + inprogress_limit, dirty: index_part.clone(), clean: (index_part, None), latest_files_changes_since_metadata_upload_scheduled: 0, @@ -325,6 +336,7 @@ impl UploadQueue { pub fn initialize_with_current_remote_index_part( &mut self, index_part: &IndexPart, + inprogress_limit: usize, ) -> anyhow::Result<&mut UploadQueueInitialized> { match self { UploadQueue::Uninitialized => (), @@ -339,6 +351,7 @@ impl UploadQueue { ); let state = UploadQueueInitialized { + inprogress_limit, dirty: index_part.clone(), clean: (index_part.clone(), None), latest_files_changes_since_metadata_upload_scheduled: 0, @@ -633,7 +646,7 @@ mod tests { #[test] fn schedule_barrier() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?; let tli = make_timeline(); let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter @@ -700,7 +713,7 @@ mod tests { #[test] fn schedule_delete_parallel() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?; let tli = make_timeline(); // Enqueue a bunch of deletes, some with conflicting names. @@ -745,7 +758,7 @@ mod tests { #[test] fn schedule_upload_conflicts() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Enqueue three versions of the same layer, with different file sizes. @@ -778,7 +791,7 @@ mod tests { #[test] fn schedule_upload_delete_conflicts() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Enqueue two layer uploads, with a delete of both layers in between them. These should be @@ -817,7 +830,7 @@ mod tests { #[test] fn schedule_upload_delete_conflicts_bypass() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Enqueue two layer uploads, with a delete of both layers in between them. These should be @@ -859,7 +872,7 @@ mod tests { #[test] fn schedule_upload_parallel() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Enqueue three different layer uploads. @@ -888,7 +901,7 @@ mod tests { #[test] fn schedule_index_serial() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; // Enqueue three uploads of the current empty index. let index = Box::new(queue.clean.0.clone()); @@ -925,7 +938,7 @@ mod tests { #[test] fn schedule_index_upload_chain() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Enqueue three uploads of the current empty index. @@ -994,7 +1007,7 @@ mod tests { #[test] fn schedule_index_delete_dereferenced() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Create a layer to upload. @@ -1038,7 +1051,7 @@ mod tests { #[test] fn schedule_index_upload_dereferenced() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example())?; + let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; let tli = make_timeline(); // Create a layer to upload. @@ -1085,7 +1098,7 @@ mod tests { #[test] fn schedule_shutdown() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; - let queue = queue.initialize_empty_remote(&TimelineMetadata::example())?; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?; let tli = make_timeline(); let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter @@ -1139,6 +1152,42 @@ mod tests { Ok(()) } + /// Scheduling respects inprogress_limit. + #[test] + fn schedule_inprogress_limit() -> anyhow::Result<()> { + // Create a queue with inprogress_limit=2. + let mut queue = UploadQueue::Uninitialized; + let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 2)?; + let tli = make_timeline(); + + // Enqueue a bunch of uploads. + let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51"); + + let ops = [ + UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None), + UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None), + UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None), + UploadOp::UploadLayer(layer3.clone(), layer3.metadata(), None), + ]; + + queue.queued_operations.extend(ops.clone()); + + // Schedule all ready operations. Only 2 are scheduled. + let tasks = queue.schedule_ready(); + assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..2]); + assert!(queue.next_ready().is_none()); + + // When one completes, another is scheduled. + queue.complete(tasks[0].task_id); + let tasks = queue.schedule_ready(); + assert_same_ops(tasks.iter().map(|t| &t.op), &ops[2..3]); + + Ok(()) + } + /// Tests that can_bypass takes name, generation and shard index into account for all operations. #[test] fn can_bypass_path() -> anyhow::Result<()> { From 6debb49b87dcf2dcca53bf17ee005b936eac446c Mon Sep 17 00:00:00 2001 From: Erik Grinaker Date: Tue, 14 Jan 2025 22:10:17 +0100 Subject: [PATCH 08/31] pageserver: coalesce index uploads when possible (#10248) ## Problem With upload queue reordering in #10218, we can easily get into a situation where multiple index uploads are queued back to back, which can't be parallelized. This will happen e.g. when multiple layer flushes enqueue layer/index/layer/index/... and the layers skip the queue and are uploaded in parallel. These index uploads will incur serial S3 roundtrip latencies, and may block later operations. Touches #10096. ## Summary of changes When multiple back-to-back index uploads are ready to upload, only upload the most recent index and drop the rest. --- pageserver/benches/upload_queue.rs | 1 + .../src/tenant/remote_timeline_client.rs | 6 +- pageserver/src/tenant/upload_queue.rs | 86 +++++++++++++------ 3 files changed, 67 insertions(+), 26 deletions(-) diff --git a/pageserver/benches/upload_queue.rs b/pageserver/benches/upload_queue.rs index ed644b0e3c..ed5daa8ae1 100644 --- a/pageserver/benches/upload_queue.rs +++ b/pageserver/benches/upload_queue.rs @@ -67,6 +67,7 @@ fn bench_upload_queue_next_ready(c: &mut Criterion) { task_id, retries: AtomicU32::new(0), op: delete.clone(), + coalesced_ops: Vec::new(), }), ); } diff --git a/pageserver/src/tenant/remote_timeline_client.rs b/pageserver/src/tenant/remote_timeline_client.rs index 1602765585..47c4a8637d 100644 --- a/pageserver/src/tenant/remote_timeline_client.rs +++ b/pageserver/src/tenant/remote_timeline_client.rs @@ -1880,7 +1880,7 @@ impl RemoteTimelineClient { /// This can launch an unbounded number of queued tasks. `UploadQueue::next_ready()` also has /// worst-case quadratic cost in the number of tasks, and may struggle beyond 10,000 tasks. fn launch_queued_tasks(self: &Arc, upload_queue: &mut UploadQueueInitialized) { - while let Some(mut next_op) = upload_queue.next_ready() { + while let Some((mut next_op, coalesced_ops)) = upload_queue.next_ready() { debug!("starting op: {next_op}"); // Prepare upload. @@ -1918,6 +1918,7 @@ impl RemoteTimelineClient { let task = Arc::new(UploadTask { task_id: upload_task_id, op: next_op, + coalesced_ops, retries: AtomicU32::new(0), }); upload_queue @@ -2285,6 +2286,9 @@ impl RemoteTimelineClient { } self.metric_end(&task.op); + for coalesced_op in &task.coalesced_ops { + self.metric_end(coalesced_op); + } } fn metric_impl( diff --git a/pageserver/src/tenant/upload_queue.rs b/pageserver/src/tenant/upload_queue.rs index 09c8f6ad8c..d302205ffe 100644 --- a/pageserver/src/tenant/upload_queue.rs +++ b/pageserver/src/tenant/upload_queue.rs @@ -22,6 +22,11 @@ use tracing::info; static DISABLE_UPLOAD_QUEUE_REORDERING: Lazy = Lazy::new(|| std::env::var("DISABLE_UPLOAD_QUEUE_REORDERING").as_deref() == Ok("true")); +/// Kill switch for index upload coalescing in case it causes problems. +/// TODO: remove this once we have confidence in it. +static DISABLE_UPLOAD_QUEUE_INDEX_COALESCING: Lazy = + Lazy::new(|| std::env::var("DISABLE_UPLOAD_QUEUE_INDEX_COALESCING").as_deref() == Ok("true")); + // clippy warns that Uninitialized is much smaller than Initialized, which wastes // memory for Uninitialized variants. Doesn't matter in practice, there are not // that many upload queues in a running pageserver, and most of them are initialized @@ -130,10 +135,12 @@ impl UploadQueueInitialized { /// the first operation in the queue, to avoid head-of-line blocking -- an operation can jump /// the queue if it doesn't conflict with operations ahead of it. /// + /// Also returns any operations that were coalesced into this one, e.g. multiple index uploads. + /// /// None may be returned even if the queue isn't empty, if no operations are ready yet. /// /// NB: this is quadratic, but queues are expected to be small, and bounded by inprogress_limit. - pub fn next_ready(&mut self) -> Option { + pub fn next_ready(&mut self) -> Option<(UploadOp, Vec)> { // If inprogress_tasks is already at limit, don't schedule anything more. if self.inprogress_limit > 0 && self.inprogress_tasks.len() >= self.inprogress_limit { return None; @@ -151,7 +158,36 @@ impl UploadQueueInitialized { return None; } - return self.queued_operations.remove(i); + let mut op = self.queued_operations.remove(i).expect("i can't disappear"); + + // Coalesce any back-to-back index uploads by only uploading the newest one that's + // ready. This typically happens with layer/index/layer/index/... sequences, where + // the layers bypass the indexes, leaving the indexes queued. + // + // If other operations are interleaved between index uploads we don't try to + // coalesce them, since we may as well update the index concurrently with them. + // This keeps the index fresh and avoids starvation. + // + // NB: we assume that all uploaded indexes have the same remote path. This + // is true at the time of writing: the path only depends on the tenant, + // timeline and generation, all of which are static for a timeline instance. + // Otherwise, we must be careful not to coalesce different paths. + let mut coalesced_ops = Vec::new(); + if matches!(op, UploadOp::UploadMetadata { .. }) { + while let Some(UploadOp::UploadMetadata { .. }) = self.queued_operations.get(i) + { + if *DISABLE_UPLOAD_QUEUE_INDEX_COALESCING { + break; + } + if !self.is_ready(i) { + break; + } + coalesced_ops.push(op); + op = self.queued_operations.remove(i).expect("i can't disappear"); + } + } + + return Some((op, coalesced_ops)); } // Nothing can bypass a barrier or shutdown. If it wasn't scheduled above, give up. @@ -225,11 +261,12 @@ impl UploadQueueInitialized { fn schedule_ready(&mut self) -> Vec> { let mut tasks = Vec::new(); // NB: schedule operations one by one, to handle conflicts with inprogress_tasks. - while let Some(op) = self.next_ready() { + while let Some((op, coalesced_ops)) = self.next_ready() { self.task_counter += 1; let task = Arc::new(UploadTask { task_id: self.task_counter, op, + coalesced_ops, retries: 0.into(), }); self.inprogress_tasks.insert(task.task_id, task.clone()); @@ -407,9 +444,13 @@ impl UploadQueue { pub struct UploadTask { /// Unique ID of this task. Used as the key in `inprogress_tasks` above. pub task_id: u64, + /// Number of task retries. pub retries: AtomicU32, - + /// The upload operation. pub op: UploadOp, + /// Any upload operations that were coalesced into this operation. This typically happens with + /// back-to-back index uploads, see `UploadQueueInitialized::next_ready()`. + pub coalesced_ops: Vec, } /// A deletion of some layers within the lifetime of a timeline. This is not used @@ -512,9 +553,8 @@ impl UploadOp { }) } - // Indexes can never bypass each other. - // TODO: we could coalesce them though, by only uploading the newest ready index. This - // is left for later, out of caution. + // Indexes can never bypass each other. They can coalesce though, and + // `UploadQueue::next_ready()` currently does this when possible. (UploadOp::UploadMetadata { .. }, UploadOp::UploadMetadata { .. }) => false, } } @@ -897,9 +937,9 @@ mod tests { Ok(()) } - /// Index uploads are serialized. + /// Index uploads are coalesced. #[test] - fn schedule_index_serial() -> anyhow::Result<()> { + fn schedule_index_coalesce() -> anyhow::Result<()> { let mut queue = UploadQueue::Uninitialized; let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?; @@ -920,13 +960,11 @@ mod tests { queue.queued_operations.extend(ops.clone()); - // The uploads should run serially. - for op in ops { - let tasks = queue.schedule_ready(); - assert_eq!(tasks.len(), 1); - assert_same_op(&tasks[0].op, &op); - queue.complete(tasks[0].task_id); - } + // The index uploads are coalesced into a single operation. + let tasks = queue.schedule_ready(); + assert_eq!(tasks.len(), 1); + assert_same_op(&tasks[0].op, &ops[2]); + assert_same_ops(&tasks[0].coalesced_ops, &ops[0..2]); assert!(queue.queued_operations.is_empty()); @@ -985,18 +1023,14 @@ mod tests { assert_same_op(&index_tasks[0].op, &ops[1]); queue.complete(index_tasks[0].task_id); - // layer 1 completes. This unblocks index 1 then index 2. + // layer 1 completes. This unblocks index 1 and 2, which coalesce into + // a single upload for index 2. queue.complete(upload_tasks[1].task_id); - let index_tasks = queue.schedule_ready(); - assert_eq!(index_tasks.len(), 1); - assert_same_op(&index_tasks[0].op, &ops[3]); - queue.complete(index_tasks[0].task_id); - let index_tasks = queue.schedule_ready(); assert_eq!(index_tasks.len(), 1); assert_same_op(&index_tasks[0].op, &ops[5]); - queue.complete(index_tasks[0].task_id); + assert_same_ops(&index_tasks[0].coalesced_ops, &ops[3..4]); assert!(queue.queued_operations.is_empty()); @@ -1018,11 +1052,12 @@ mod tests { let index_deref = index_without(&index_upload, &layer); let ops = [ - // Initial upload. + // Initial upload, with a barrier to prevent index coalescing. UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), UploadOp::UploadMetadata { uploaded: index_upload.clone(), }, + UploadOp::Barrier(tokio::sync::watch::channel(()).0), // Dereference the layer and delete it. UploadOp::UploadMetadata { uploaded: index_deref.clone(), @@ -1063,11 +1098,12 @@ mod tests { let index_ref = index_with(&index_deref, &layer); let ops = [ - // Initial upload. + // Initial upload, with a barrier to prevent index coalescing. UploadOp::UploadLayer(layer.clone(), layer.metadata(), None), UploadOp::UploadMetadata { uploaded: index_upload.clone(), }, + UploadOp::Barrier(tokio::sync::watch::channel(()).0), // Dereference the layer. UploadOp::UploadMetadata { uploaded: index_deref.clone(), From 47c1640accbbf45071de91b28aadc7356775fc43 Mon Sep 17 00:00:00 2001 From: John Spray Date: Tue, 14 Jan 2025 21:37:32 +0000 Subject: [PATCH 09/31] storage controller: pagination for tenant listing API (#10365) ## Problem For large deployments, the `control/v1/tenant` listing API can time out transmitting a monolithic serialized response. ## Summary of changes - Add `limit` and `start_after` parameters to listing API - Update storcon_cli to use these parameters and limit requests to 1000 items at a time --- control_plane/storcon_cli/src/main.rs | 74 +++++++++++++------ storage_controller/src/http.rs | 6 +- storage_controller/src/service.rs | 33 ++++++++- test_runner/fixtures/neon_fixtures.py | 19 ++++- .../regress/test_storage_controller.py | 19 ++++- 5 files changed, 117 insertions(+), 34 deletions(-) diff --git a/control_plane/storcon_cli/src/main.rs b/control_plane/storcon_cli/src/main.rs index 9d133e4af1..2ba8f63678 100644 --- a/control_plane/storcon_cli/src/main.rs +++ b/control_plane/storcon_cli/src/main.rs @@ -477,16 +477,7 @@ async fn main() -> anyhow::Result<()> { println!("{table}"); } Command::Tenants { node_id: None } => { - let mut resp = storcon_client - .dispatch::<(), Vec>( - Method::GET, - "control/v1/tenant".to_string(), - None, - ) - .await?; - - resp.sort_by(|a, b| a.tenant_id.cmp(&b.tenant_id)); - + // Set up output formatting let mut table = comfy_table::Table::new(); table.set_header([ "TenantId", @@ -496,20 +487,55 @@ async fn main() -> anyhow::Result<()> { "Placement", "Scheduling", ]); - for tenant in resp { - let shard_zero = tenant.shards.into_iter().next().unwrap(); - table.add_row([ - format!("{}", tenant.tenant_id), - shard_zero - .preferred_az_id - .as_ref() - .cloned() - .unwrap_or("".to_string()), - format!("{}", shard_zero.tenant_shard_id.shard_count.literal()), - format!("{:?}", tenant.stripe_size), - format!("{:?}", tenant.policy), - format!("{:?}", shard_zero.scheduling_policy), - ]); + + // Pagination loop over listing API + let mut start_after = None; + const LIMIT: usize = 1000; + loop { + let path = match start_after { + None => format!("control/v1/tenant?limit={LIMIT}"), + Some(start_after) => { + format!("control/v1/tenant?limit={LIMIT}&start_after={start_after}") + } + }; + + let resp = storcon_client + .dispatch::<(), Vec>(Method::GET, path, None) + .await?; + + if resp.is_empty() { + // End of data reached + break; + } + + // Give some visual feedback while we're building up the table (comfy_table doesn't have + // streaming output) + if resp.len() >= LIMIT { + eprint!("."); + } + + start_after = Some(resp.last().unwrap().tenant_id); + + for tenant in resp { + let shard_zero = tenant.shards.into_iter().next().unwrap(); + table.add_row([ + format!("{}", tenant.tenant_id), + shard_zero + .preferred_az_id + .as_ref() + .cloned() + .unwrap_or("".to_string()), + format!("{}", shard_zero.tenant_shard_id.shard_count.literal()), + format!("{:?}", tenant.stripe_size), + format!("{:?}", tenant.policy), + format!("{:?}", shard_zero.scheduling_policy), + ]); + } + } + + // Terminate progress dots + if table.row_count() > LIMIT { + eprint!(""); } println!("{table}"); diff --git a/storage_controller/src/http.rs b/storage_controller/src/http.rs index c8df4ffe28..03d8f11992 100644 --- a/storage_controller/src/http.rs +++ b/storage_controller/src/http.rs @@ -653,6 +653,10 @@ async fn handle_tenant_list( ) -> Result, ApiError> { check_permissions(&req, Scope::Admin)?; + let limit: Option = parse_query_param(&req, "limit")?; + let start_after: Option = parse_query_param(&req, "start_after")?; + tracing::info!("start_after: {:?}", start_after); + match maybe_forward(req).await { ForwardOutcome::Forwarded(res) => { return res; @@ -660,7 +664,7 @@ async fn handle_tenant_list( ForwardOutcome::NotForwarded(_req) => {} }; - json_response(StatusCode::OK, service.tenant_list()) + json_response(StatusCode::OK, service.tenant_list(limit, start_after)) } async fn handle_node_register(req: Request) -> Result, ApiError> { diff --git a/storage_controller/src/service.rs b/storage_controller/src/service.rs index cbb9103880..57f4cc8463 100644 --- a/storage_controller/src/service.rs +++ b/storage_controller/src/service.rs @@ -4158,17 +4158,42 @@ impl Service { .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into())) } - pub(crate) fn tenant_list(&self) -> Vec { + /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not + /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory + /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses + /// in our external API. + pub(crate) fn tenant_list( + &self, + limit: Option, + start_after: Option, + ) -> Vec { let locked = self.inner.read().unwrap(); + // Apply start_from parameter + let shard_range = match start_after { + None => locked.tenants.range(..), + Some(tenant_id) => locked.tenants.range( + TenantShardId { + tenant_id, + shard_number: ShardNumber(u8::MAX), + shard_count: ShardCount(u8::MAX), + }.., + ), + }; + let mut result = Vec::new(); - for (_tenant_id, tenant_shards) in - &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id) - { + for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) { result.push( self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v)) .expect("Groups are always non-empty"), ); + + // Enforce `limit` parameter + if let Some(limit) = limit { + if result.len() >= limit { + break; + } + } } result diff --git a/test_runner/fixtures/neon_fixtures.py b/test_runner/fixtures/neon_fixtures.py index e22e452a52..c47739cd81 100644 --- a/test_runner/fixtures/neon_fixtures.py +++ b/test_runner/fixtures/neon_fixtures.py @@ -1884,7 +1884,10 @@ class NeonStorageController(MetricsGetter, LogUtils): ) return response.json() - def tenant_list(self): + def tenant_shard_dump(self): + """ + Debug listing API: dumps the internal map of tenant shards + """ response = self.request( "GET", f"{self.api}/debug/v1/tenant", @@ -1892,6 +1895,18 @@ class NeonStorageController(MetricsGetter, LogUtils): ) return response.json() + def tenant_list(self, **kwargs): + """ + Control API tenant listing: a vector of the same content returned by tenant_describe + """ + response = self.request( + "GET", + f"{self.api}/control/v1/tenant", + headers=self.headers(TokenScope.ADMIN), + params=kwargs, + ) + return response.json() + def node_configure(self, node_id, body: dict[str, Any]): log.info(f"node_configure({node_id}, {body})") body["node_id"] = node_id @@ -2238,7 +2253,7 @@ class NeonStorageController(MetricsGetter, LogUtils): """ Get the intent and observed placements of all tenants known to the storage controller. """ - tenants = self.tenant_list() + tenants = self.tenant_shard_dump() tenant_placement: defaultdict[str, dict[str, Any]] = defaultdict( lambda: { diff --git a/test_runner/regress/test_storage_controller.py b/test_runner/regress/test_storage_controller.py index 8ffb6ba6b2..b5d109559f 100644 --- a/test_runner/regress/test_storage_controller.py +++ b/test_runner/regress/test_storage_controller.py @@ -113,6 +113,19 @@ def test_storage_controller_smoke(neon_env_builder: NeonEnvBuilder, combination) for tid in tenant_ids: env.create_tenant(tid, shard_count=shards_per_tenant) + # Tenant listing API should work + listed_tenants = env.storage_controller.tenant_list() + log.info(f"listed_tenants: {listed_tenants}") + assert set(t["tenant_id"] for t in listed_tenants) == set(str(t) for t in tenant_ids) + paged = env.storage_controller.tenant_list(limit=2, start_after=listed_tenants[0]["tenant_id"]) + assert len(paged) == 2 + assert paged[0] == listed_tenants[1] + assert paged[1] == listed_tenants[2] + paged = env.storage_controller.tenant_list( + limit=1000, start_after="ffffffffffffffffffffffffffffffff" + ) + assert paged == [] + # Validate high level metrics assert ( env.storage_controller.get_metric_value("storage_controller_tenant_shards") @@ -1506,7 +1519,7 @@ class PageserverFailpoint(Failure): def build_node_to_tenants_map(env: NeonEnv) -> dict[int, list[TenantId]]: - tenants = env.storage_controller.tenant_list() + tenants = env.storage_controller.tenant_shard_dump() node_to_tenants: dict[int, list[TenantId]] = {} for t in tenants: @@ -2631,7 +2644,7 @@ def test_storage_controller_step_down(neon_env_builder: NeonEnvBuilder): # Validate that the storcon attempts to forward the request, but stops. # when it realises it is still the current leader. with pytest.raises(StorageControllerApiException, match="Leader is stepped down instance"): - env.storage_controller.tenant_list() + env.storage_controller.tenant_shard_dump() # Validate that we can step down multiple times and the observed state # doesn't change. @@ -2781,7 +2794,7 @@ def test_storage_controller_leadership_transfer( # Check that the stepped down instance forwards requests # to the new leader while it's still running. storage_controller_proxy.route_to(f"http://127.0.0.1:{storage_controller_1_port}") - env.storage_controller.tenant_list() + env.storage_controller.tenant_shard_dump() env.storage_controller.node_configure(env.pageservers[0].id, {"scheduling": "Pause"}) status = env.storage_controller.node_status(env.pageservers[0].id) assert status["scheduling"] == "Pause" From c98cbbeac143837eb99dfea1b5246bad21647b22 Mon Sep 17 00:00:00 2001 From: Arseny Sher Date: Wed, 15 Jan 2025 12:41:49 +0300 Subject: [PATCH 10/31] Add migration details to safekeeper membership RFC. (#10272) ## Problem https://github.com/neondatabase/neon/pull/8455 wasn't specific enough on migration from current situation to enabling generations. ## Summary of changes Describe the missing parts, including control plane pushing generation to compute, which also defines whether generations are enabled -- non zero value does it. --- ...35-safekeeper-dynamic-membership-change.md | 188 ++++++++++++------ 1 file changed, 122 insertions(+), 66 deletions(-) diff --git a/docs/rfcs/035-safekeeper-dynamic-membership-change.md b/docs/rfcs/035-safekeeper-dynamic-membership-change.md index 239ec58186..cea9af34ab 100644 --- a/docs/rfcs/035-safekeeper-dynamic-membership-change.md +++ b/docs/rfcs/035-safekeeper-dynamic-membership-change.md @@ -81,7 +81,7 @@ configuration generation in them is less than its current one. Namely, it refuses to vote, to truncate WAL in `handle_elected` and to accept WAL. In response it sends its current configuration generation to let walproposer know. -Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration` +Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration` accepting `Configuration`. Safekeeper switches to the given conf it is higher than its current one and ignores it otherwise. In any case it replies with ``` @@ -103,7 +103,7 @@ currently and tries to communicate with all of them. However, the list does not define consensus members. Instead, on start walproposer tracks highest configuration it receives from `AcceptorGreeting`s. Once it assembles greetings from majority of `sk_set` and majority of `new_sk_set` (if it is present), it -establishes this configuration as its own and moves to voting. +establishes this configuration as its own and moves to voting. It should stop talking to safekeepers not listed in the configuration at this point, though it is not unsafe to continue doing so. @@ -119,7 +119,7 @@ refusal to accept due to configuration change) it simply restarts. The following algorithm can be executed anywhere having access to configuration storage and safekeepers. It is safe to interrupt / restart it and run multiple instances of it concurrently, though likely one of them won't make -progress then. It accepts `desired_set: Vec` as input. +progress then. It accepts `desired_set: Vec` as input. Algorithm will refuse to make the change if it encounters previous interrupted change attempt, but in this case it will try to finish it. @@ -140,7 +140,7 @@ storage are reachable. safe. Failed CAS aborts the procedure. 4) Call `PUT` `configuration` on safekeepers from the current set, delivering them `joint_conf`. Collecting responses from majority is required - to proceed. If any response returned generation higher than + to proceed. If any response returned generation higher than `joint_conf.generation`, abort (another switch raced us). Otherwise, choose max `` among responses and establish it as (in memory) `sync_position`. Also choose max `term` and establish it as (in @@ -149,49 +149,49 @@ storage are reachable. without ack from the new set. Similarly, we'll bump term on new majority to `sync_term` so that two computes with the same term are never elected. 4) Initialize timeline on safekeeper(s) from `new_sk_set` where it - doesn't exist yet by doing `pull_timeline` from the majority of the + doesn't exist yet by doing `pull_timeline` from the majority of the current set. Doing that on majority of `new_sk_set` is enough to proceed, but it is reasonable to ensure that all `new_sk_set` members are initialized -- if some of them are down why are we migrating there? -5) Call `POST` `bump_term(sync_term)` on safekeepers from the new set. +5) Call `POST` `bump_term(sync_term)` on safekeepers from the new set. Success on majority is enough. 6) Repeatedly call `PUT` `configuration` on safekeepers from the new set, delivering them `joint_conf` and collecting their positions. This will - switch them to the `joint_conf` which generally won't be needed + switch them to the `joint_conf` which generally won't be needed because `pull_timeline` already includes it and plus additionally would be broadcast by compute. More importantly, we may proceed to the next step - only when `` on the majority of the new set reached - `sync_position`. Similarly, on the happy path no waiting is not needed because + only when `` on the majority of the new set reached + `sync_position`. Similarly, on the happy path no waiting is not needed because `pull_timeline` already includes it. However, we should double check to be safe. For example, timeline could have been created earlier e.g. - manually or after try-to-migrate, abort, try-to-migrate-again sequence. -7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new - safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration + manually or after try-to-migrate, abort, try-to-migrate-again sequence. +7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new + safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration storage under one more CAS. 8) Call `PUT` `configuration` on safekeepers from the new set, - delivering them `new_conf`. It is enough to deliver it to the majority + delivering them `new_conf`. It is enough to deliver it to the majority of the new set; the rest can be updated by compute. I haven't put huge effort to make the description above very precise, because it is natural language prone to interpretations anyway. Instead I'd like to make TLA+ spec of it. -Description above focuses on safety. To make the flow practical and live, here a few more +Description above focuses on safety. To make the flow practical and live, here a few more considerations. -1) It makes sense to ping new set to ensure it we are migrating to live node(s) before +1) It makes sense to ping new set to ensure it we are migrating to live node(s) before step 3. -2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed +2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed it is safe to rollback to the old conf with one more CAS. -3) On step 4 timeline might be already created on members of the new set for various reasons; +3) On step 4 timeline might be already created on members of the new set for various reasons; the simplest is the procedure restart. There are more complicated scenarious like mentioned - in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving - generations, so seems simpler to treat existing timeline as success. However, this also + in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving + generations, so seems simpler to treat existing timeline as success. However, this also has a disadvantage: you might imagine an surpassingly unlikely schedule where condition in the step 5 is never reached until compute is (re)awaken up to synchronize new member(s). I don't think we'll observe this in practice, but can add waking up compute if needed. 4) In the end timeline should be locally deleted on the safekeeper(s) which are in the old set but not in the new one, unless they are unreachable. To be - safe this also should be done under generation number (deletion proceeds only if + safe this also should be done under generation number (deletion proceeds only if current configuration is <= than one in request and safekeeper is not memeber of it). 5) If current conf fetched on step 1 is already not joint and members equal to `desired_set`, jump to step 7, using it as `new_conf`. @@ -202,47 +202,87 @@ The procedure ought to be driven from somewhere. Obvious candidates are control plane and storage_controller; and as each of them already has db we don't want yet another storage. I propose to manage safekeepers in storage_controller because 1) since it is in rust it simplifies simulation testing (more on this -below) 2) it already manages pageservers. +below) 2) it already manages pageservers. This assumes that migration will be fully usable only after we migrate all tenants/timelines to storage_controller. It is discussible whether we want also to manage pageserver attachments for all of these, but likely we do. -This requires us to define storcon <-> cplane interface. +This requires us to define storcon <-> cplane interface and changes. -### storage_controller <-> control plane interface +### storage_controller <-> control plane interface and changes First of all, control plane should [change](https://neondb.slack.com/archives/C03438W3FLZ/p1719226543199829) storing safekeepers per timeline instead of per tenant because we can't migrate -tenants atomically. +tenants atomically. The important question is how updated configuration is delivered from storage_controller to control plane to provide it to computes. As always, there are two options, pull and push. Let's do it the same push as with pageserver `/notify-attach` because 1) it keeps storage_controller out of critical compute -start path 2) provides easier upgrade: there won't be such a thing as 'timeline -managed by control plane / storcon', cplane just takes the value out of its db -when needed 3) uniformity. It makes storage_controller responsible for retrying notifying -control plane until it succeeds. +start path 2) uniformity. It makes storage_controller responsible for retrying +notifying control plane until it succeeds. -So, cplane `/notify-safekeepers` for the timeline accepts `Configuration` and -updates it in the db if the provided conf generation is higher (the cplane db -should also store generations for this). Similarly to [`/notify-attach`](https://www.notion.so/neondatabase/Storage-Controller-Control-Plane-interface-6de56dd310a043bfa5c2f5564fa98365), it -should update db which makes the call successful, and then try to schedule -`apply_config` if possible, it is ok if not. storage_controller -should rate limit calling the endpoint, but likely this won't be needed, as migration +It is not needed for the control plane to fully know the `Configuration`. It is +enough for it to only to be aware of the list of safekeepers in the latest +configuration to supply it to compute, plus associated generation number to +protect from stale update requests and to also pass it to compute. + +So, cplane `/notify-safekeepers` for the timeline can accept JSON like +``` +{ + tenant_id: String, + timeline_id: String, + generation: u32, + safekeepers: Vec, +} +``` +where `SafekeeperId` is +``` +{ + node_id: u64, + host: String +} +``` +In principle `host` is redundant, but may be useful for observability. + +The request updates list of safekeepers in the db if the provided conf +generation is higher (the cplane db should also store generations for this). +Similarly to +[`/notify-attach`](https://www.notion.so/neondatabase/Storage-Controller-Control-Plane-interface-6de56dd310a043bfa5c2f5564fa98365), +it should update db which makes the call successful, and then try to schedule +`apply_config` if possible, it is ok if not. storage_controller should rate +limit calling the endpoint, but likely this won't be needed, as migration throughput is limited by `pull_timeline`. Timeline (branch) creation in cplane should call storage_controller POST `tenant/:tenant_id/timeline` like it currently does for sharded tenants. -Response should be augmented with `safekeeper_conf: Configuration`. The call -should be retried until succeeds. +Response should be augmented with `safekeepers_generation` and `safekeepers` +fields like described in `/notify-safekeepers` above. Initially (currently) +these fields may be absent; in this case cplane chooses safekeepers on its own +like it currently does. The call should be retried until succeeds. Timeline deletion and tenant deletion in cplane should call appropriate storage_controller endpoints like it currently does for sharded tenants. The calls should be retried until they succeed. +When compute receives safekeepers list from control plane it needs to know the +generation to checked whether it should be updated (note that compute may get +safekeeper list from either cplane or safekeepers). Currently `neon.safekeepers` +GUC is just a comma separates list of `host:port`. Let's prefix it with +`g#:` to this end, so it will look like +``` +g#42:safekeeper-0.eu-central-1.aws.neon.tech:6401,safekeeper-2.eu-central-1.aws.neon.tech:6401,safekeeper-1.eu-central-1.aws.neon.tech:6401 +``` + +To summarize, list of cplane changes: +- per tenant -> per timeline safekeepers management and addition of int `safekeeper_generation` field. +- `/notify-safekeepers` endpoint. +- Branch creation call may return list of safekeepers and when it is + present cplane should adopt it instead of choosing on its own like it does currently. +- `neon.safekeepers` GUC should be prefixed with `g#:`. + ### storage_controller implementation Current 'load everything on startup and keep in memory' easy design is fine. @@ -360,10 +400,10 @@ source safekeeper might fail, which is not a problem if we are going to decomission the node but leaves garbage otherwise. I'd propose in the first version 1) Don't attempt deletion at all if node status is `offline`. 2) If it failed, just issue warning. -And add PUT `/control/v1/safekeepers/:node_id/scrub` endpoint which would find and -remove garbage timelines for manual use. It will 1) list all timelines on the -safekeeper 2) compare each one against configuration storage: if timeline -doesn't exist at all (had been deleted), it can be deleted. Otherwise, it can +And add PUT `/control/v1/safekeepers/:node_id/scrub` endpoint which would find and +remove garbage timelines for manual use. It will 1) list all timelines on the +safekeeper 2) compare each one against configuration storage: if timeline +doesn't exist at all (had been deleted), it can be deleted. Otherwise, it can be deleted under generation number if node is not member of current generation. Automating this is untrivial; we'd need to register all potential missing @@ -412,8 +452,8 @@ There should be following layers of tests: 3) Since simulation testing injects at relatively high level points (not syscalls), it omits some code, in particular `pull_timeline`. Thus it is better to have basic tests covering whole system as well. Extended version of - `test_restarts_under_load` would do: start background load and do migration - under it, then restart endpoint and check that no reported commits + `test_restarts_under_load` would do: start background load and do migration + under it, then restart endpoint and check that no reported commits had been lost. I'd also add one more creating classic network split scenario, with one compute talking to AC and another to BD while migration from nodes ABC to ABD happens. @@ -422,35 +462,51 @@ There should be following layers of tests: ## Order of implementation and rollout -Note that +Note that - Control plane parts and integration with it is fully independent from everything else (tests would use simulation and neon_local). +- It is reasonable to make compute <-> safekeepers protocol change + independent of enabling generations. - There is a lot of infra work making storage_controller aware of timelines and safekeepers and its impl/rollout should be separate from migration itself. -- Initially walproposer can just stop working while it observers joint configuration. +- Initially walproposer can just stop working while it observes joint configuration. Such window would be typically very short anyway. +- Obviously we want to test the whole thing thoroughly on staging and only then + gradually enable in prod. -To rollout smoothly, both walproposer and safekeeper should have flag -`configurations_enabled`; when set to false, they would work as currently, i.e. -walproposer is able to commit on whatever safekeeper set it is provided. Until -all timelines are managed by storcon we'd need to use current script to migrate -and update/drop entries in the storage_controller database if it has any. +Let's have the following implementation bits for gradual rollout: +- compute gets `neon.safekeepers_proto_version` flag. + Initially both compute and safekeepers will be able to talk both + versions so that we can delay force restart of them and for + simplicity of rollback in case it is needed. +- storcon gets `-set-safekeepers` config option disabled by + default. Timeline creation request chooses safekeepers + (and returns them in response to cplane) only when it is set to + true. +- control_plane [see above](storage_controller-<->-control-plane interface-and-changes) + prefixes `neon.safekeepers` GUC with generation number. When it is 0 + (or prefix not present at all), walproposer behaves as currently, committing on + the provided safekeeper list -- generations are disabled. + If it is non 0 it follows this RFC rules. +- We provide a script for manual migration to storage controller. + It selects timeline(s) from control plane (specified or all of them) db + and calls special import endpoint on storage controller which is very + similar to timeline creation: it inserts into the db, sets + configuration to initial on the safekeepers, calls cplane + `notify-safekeepers`. -Safekeepers would need to be able to talk both current and new protocol version -with compute to reduce number of computes restarted in prod once v2 protocol is -deployed (though before completely switching we'd need to force this). - -Let's have the following rollout order: -- storage_controller becomes aware of safekeepers; -- storage_controller gets timeline creation for new timelines and deletion requests, but - doesn't manage all timelines yet. Migration can be tested on these new timelines. - To keep control plane and storage_controller databases in sync while control - plane still chooses the safekeepers initially (until all timelines are imported - it can choose better), `TimelineCreateRequest` can get optional safekeepers - field with safekeepers chosen by cplane. -- Then we can import all existing timelines from control plane to - storage_controller and gradually enable configurations region by region. +Then the rollout for a region would be: +- Current situation: safekeepers are choosen by control_plane. +- We manually migrate some timelines, test moving them around. +- Then we enable `--set-safekeepers` so that all new timelines + are on storage controller. +- Finally migrate all existing timelines using the script (no + compute should be speaking old proto version at this point). +Until all timelines are managed by storcon we'd need to use current ad hoc +script to migrate if needed. To keep state clean, all storage controller managed +timelines must be migrated before that, or controller db and configurations +state of safekeepers dropped manually. Very rough implementation order: - Add concept of configurations to safekeepers (including control file), @@ -458,10 +514,10 @@ Very rough implementation order: - Implement walproposer changes, including protocol. - Implement storconn part. Use it in neon_local (and pytest). - Make cplane store safekeepers per timeline instead of per tenant. -- Implement cplane/storcon integration. Route branch creation/deletion +- Implement cplane/storcon integration. Route branch creation/deletion through storcon. Then we can test migration of new branches. -- Finally import existing branches. Then we can drop cplane - safekeeper selection code. Gradually enable configurations at +- Finally import existing branches. Then we can drop cplane + safekeeper selection code. Gradually enable configurations at computes and safekeepers. Before that, all computes must talk only v3 protocol version. From 2d0ea085244208a8b0238b69c7dc5a4f93890d90 Mon Sep 17 00:00:00 2001 From: Arseny Sher Date: Wed, 15 Jan 2025 12:45:58 +0300 Subject: [PATCH 11/31] Add safekeeper membership conf to control file. (#10196) ## Problem https://github.com/neondatabase/neon/issues/9965 ## Summary of changes Add safekeeper membership configuration struct itself and storing it in the control file. In passing also add creation timestamp to the control file (there were cases where I wanted it in the past). Remove obsolete unused PersistedPeerInfo struct from control file (still keep it control_file_upgrade.rs to have it in old upgrade code). Remove the binary representation of cfile in the roundtrip test. Updating it is annoying, and we still test the actual roundtrip. Also add configuration to timeline creation http request, currently used only in one python test. In passing, slightly change LSNs meaning in the request: normally start_lsn is passed (the same as ancestor_start_lsn in similar pageserver call), but we allow specifying higher commit_lsn for manual intervention if needed. Also when given LSN initialize term_history with it. --- Cargo.lock | 2 + libs/safekeeper_api/Cargo.toml | 2 + libs/safekeeper_api/src/lib.rs | 5 +- libs/safekeeper_api/src/membership.rs | 164 ++++++++++++++++++ libs/safekeeper_api/src/models.rs | 13 +- safekeeper/src/control_file.rs | 20 ++- safekeeper/src/control_file_upgrade.rs | 145 ++++++++++++++-- safekeeper/src/copy_timeline.rs | 5 +- safekeeper/src/http/routes.rs | 13 +- safekeeper/src/json_ctrl.rs | 2 + safekeeper/src/receive_wal.rs | 9 +- safekeeper/src/safekeeper.rs | 117 +++---------- safekeeper/src/state.rs | 69 ++++---- safekeeper/src/timelines_global_map.rs | 7 +- .../tests/walproposer_sim/safekeeper.rs | 11 +- test_runner/fixtures/pageserver/http.py | 11 +- test_runner/fixtures/safekeeper/http.py | 46 +++-- test_runner/fixtures/utils.py | 18 ++ test_runner/regress/test_wal_acceptor.py | 10 +- 19 files changed, 477 insertions(+), 192 deletions(-) create mode 100644 libs/safekeeper_api/src/membership.rs diff --git a/Cargo.lock b/Cargo.lock index 1e29f4fc08..f543057933 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -5707,10 +5707,12 @@ dependencies = [ name = "safekeeper_api" version = "0.1.0" dependencies = [ + "anyhow", "const_format", "postgres_ffi", "pq_proto", "serde", + "serde_json", "tokio", "utils", ] diff --git a/libs/safekeeper_api/Cargo.toml b/libs/safekeeper_api/Cargo.toml index 4234ec6779..7652c3d413 100644 --- a/libs/safekeeper_api/Cargo.toml +++ b/libs/safekeeper_api/Cargo.toml @@ -5,8 +5,10 @@ edition.workspace = true license.workspace = true [dependencies] +anyhow.workspace = true const_format.workspace = true serde.workspace = true +serde_json.workspace = true postgres_ffi.workspace = true pq_proto.workspace = true tokio.workspace = true diff --git a/libs/safekeeper_api/src/lib.rs b/libs/safekeeper_api/src/lib.rs index be6923aca9..fa86523ad7 100644 --- a/libs/safekeeper_api/src/lib.rs +++ b/libs/safekeeper_api/src/lib.rs @@ -4,12 +4,15 @@ use const_format::formatcp; use pq_proto::SystemId; use serde::{Deserialize, Serialize}; +pub mod membership; /// Public API types pub mod models; /// Consensus logical timestamp. Note: it is a part of sk control file. pub type Term = u64; -pub const INVALID_TERM: Term = 0; +/// With this term timeline is created initially. It +/// is a normal term except wp is never elected with it. +pub const INITIAL_TERM: Term = 0; /// Information about Postgres. Safekeeper gets it once and then verifies all /// further connections from computes match. Note: it is a part of sk control diff --git a/libs/safekeeper_api/src/membership.rs b/libs/safekeeper_api/src/membership.rs new file mode 100644 index 0000000000..fe30204545 --- /dev/null +++ b/libs/safekeeper_api/src/membership.rs @@ -0,0 +1,164 @@ +//! Types defining safekeeper membership, see +//! rfcs/035-safekeeper-dynamic-membership-change.md +//! for details. + +use std::{collections::HashSet, fmt::Display}; + +use anyhow; +use anyhow::bail; +use serde::{Deserialize, Serialize}; +use utils::id::NodeId; + +/// Number uniquely identifying safekeeper configuration. +/// Note: it is a part of sk control file. +pub type Generation = u32; +/// 1 is the first valid generation, 0 is used as +/// a placeholder before we fully migrate to generations. +pub const INVALID_GENERATION: Generation = 0; +pub const INITIAL_GENERATION: Generation = 1; + +/// Membership is defined by ids so e.g. walproposer uses them to figure out +/// quorums, but we also carry host and port to give wp idea where to connect. +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +pub struct SafekeeperId { + pub id: NodeId, + pub host: String, + pub pg_port: u16, +} + +impl Display for SafekeeperId { + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + write!(f, "[id={}, ep={}:{}]", self.id, self.host, self.pg_port) + } +} + +/// Set of safekeepers. +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +#[serde(transparent)] +pub struct MemberSet { + pub members: Vec, +} + +impl MemberSet { + pub fn empty() -> Self { + MemberSet { + members: Vec::new(), + } + } + + pub fn new(members: Vec) -> anyhow::Result { + let hs: HashSet = HashSet::from_iter(members.iter().map(|sk| sk.id)); + if hs.len() != members.len() { + bail!("duplicate safekeeper id in the set {:?}", members); + } + Ok(MemberSet { members }) + } + + pub fn contains(&self, sk: &SafekeeperId) -> bool { + self.members.iter().any(|m| m.id == sk.id) + } + + pub fn add(&mut self, sk: SafekeeperId) -> anyhow::Result<()> { + if self.contains(&sk) { + bail!(format!( + "sk {} is already member of the set {}", + sk.id, self + )); + } + self.members.push(sk); + Ok(()) + } +} + +impl Display for MemberSet { + /// Display as a comma separated list of members. + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + let sks_str = self + .members + .iter() + .map(|m| m.to_string()) + .collect::>(); + write!(f, "({})", sks_str.join(", ")) + } +} + +/// Safekeeper membership configuration. +/// Note: it is a part of both control file and http API. +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +pub struct Configuration { + /// Unique id. + pub generation: Generation, + /// Current members of the configuration. + pub members: MemberSet, + /// Some means it is a joint conf. + pub new_members: Option, +} + +impl Configuration { + /// Used for pre-generations timelines, will be removed eventually. + pub fn empty() -> Self { + Configuration { + generation: INVALID_GENERATION, + members: MemberSet::empty(), + new_members: None, + } + } +} + +impl Display for Configuration { + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + write!( + f, + "gen={}, members={}, new_members={}", + self.generation, + self.members, + self.new_members + .as_ref() + .map(ToString::to_string) + .unwrap_or(String::from("none")) + ) + } +} + +#[cfg(test)] +mod tests { + use super::{MemberSet, SafekeeperId}; + use utils::id::NodeId; + + #[test] + fn test_member_set() { + let mut members = MemberSet::empty(); + members + .add(SafekeeperId { + id: NodeId(42), + host: String::from("lala.org"), + pg_port: 5432, + }) + .unwrap(); + + members + .add(SafekeeperId { + id: NodeId(42), + host: String::from("lala.org"), + pg_port: 5432, + }) + .expect_err("duplicate must not be allowed"); + + members + .add(SafekeeperId { + id: NodeId(43), + host: String::from("bubu.org"), + pg_port: 5432, + }) + .unwrap(); + + println!("members: {}", members); + + let j = serde_json::to_string(&members).expect("failed to serialize"); + println!("members json: {}", j); + assert_eq!( + j, + r#"[{"id":42,"host":"lala.org","pg_port":5432},{"id":43,"host":"bubu.org","pg_port":5432}]"# + ); + } +} diff --git a/libs/safekeeper_api/src/models.rs b/libs/safekeeper_api/src/models.rs index 3e424a792c..ad38986357 100644 --- a/libs/safekeeper_api/src/models.rs +++ b/libs/safekeeper_api/src/models.rs @@ -11,7 +11,7 @@ use utils::{ pageserver_feedback::PageserverFeedback, }; -use crate::{ServerInfo, Term}; +use crate::{membership::Configuration, ServerInfo, Term}; #[derive(Debug, Serialize)] pub struct SafekeeperStatus { @@ -22,13 +22,16 @@ pub struct SafekeeperStatus { pub struct TimelineCreateRequest { pub tenant_id: TenantId, pub timeline_id: TimelineId, - pub peer_ids: Option>, + pub mconf: Configuration, pub pg_version: u32, pub system_id: Option, + // By default WAL_SEGMENT_SIZE pub wal_seg_size: Option, - pub commit_lsn: Lsn, - // If not passed, it is assigned to the beginning of commit_lsn segment. - pub local_start_lsn: Option, + pub start_lsn: Lsn, + // Normal creation should omit this field (start_lsn initializes all LSNs). + // However, we allow specifying custom value higher than start_lsn for + // manual recovery case, see test_s3_wal_replay. + pub commit_lsn: Option, } /// Same as TermLsn, but serializes LSN using display serializer diff --git a/safekeeper/src/control_file.rs b/safekeeper/src/control_file.rs index 06e5afbf74..e92ca881e1 100644 --- a/safekeeper/src/control_file.rs +++ b/safekeeper/src/control_file.rs @@ -3,6 +3,7 @@ use anyhow::{bail, ensure, Context, Result}; use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt}; use camino::{Utf8Path, Utf8PathBuf}; +use safekeeper_api::membership::INVALID_GENERATION; use tokio::fs::File; use tokio::io::AsyncWriteExt; use utils::crashsafe::durable_rename; @@ -13,14 +14,14 @@ use std::ops::Deref; use std::path::Path; use std::time::Instant; -use crate::control_file_upgrade::downgrade_v9_to_v8; +use crate::control_file_upgrade::downgrade_v10_to_v9; use crate::control_file_upgrade::upgrade_control_file; use crate::metrics::PERSIST_CONTROL_FILE_SECONDS; use crate::state::{EvictionState, TimelinePersistentState}; use utils::bin_ser::LeSer; pub const SK_MAGIC: u32 = 0xcafeceefu32; -pub const SK_FORMAT_VERSION: u32 = 9; +pub const SK_FORMAT_VERSION: u32 = 10; // contains persistent metadata for safekeeper pub const CONTROL_FILE_NAME: &str = "safekeeper.control"; @@ -169,10 +170,11 @@ impl TimelinePersistentState { let mut buf: Vec = Vec::new(); WriteBytesExt::write_u32::(&mut buf, SK_MAGIC)?; - if self.eviction_state == EvictionState::Present { - // temp hack for forward compatibility - const PREV_FORMAT_VERSION: u32 = 8; - let prev = downgrade_v9_to_v8(self); + if self.mconf.generation == INVALID_GENERATION { + // Temp hack for forward compatibility test: in case of none + // configuration save cfile in previous v9 format. + const PREV_FORMAT_VERSION: u32 = 9; + let prev = downgrade_v10_to_v9(self); WriteBytesExt::write_u32::(&mut buf, PREV_FORMAT_VERSION)?; prev.ser_into(&mut buf)?; } else { @@ -233,6 +235,7 @@ impl Storage for FileStorage { #[cfg(test)] mod test { use super::*; + use safekeeper_api::membership::{Configuration, MemberSet}; use tokio::fs; use utils::lsn::Lsn; @@ -242,6 +245,11 @@ mod test { async fn test_read_write_safekeeper_state() -> anyhow::Result<()> { let tempdir = camino_tempfile::tempdir()?; let mut state = TimelinePersistentState::empty(); + state.mconf = Configuration { + generation: 42, + members: MemberSet::empty(), + new_members: None, + }; let mut storage = FileStorage::create_new(tempdir.path(), state.clone(), NO_SYNC).await?; // Make a change. diff --git a/safekeeper/src/control_file_upgrade.rs b/safekeeper/src/control_file_upgrade.rs index dd152fd4cc..904e79f976 100644 --- a/safekeeper/src/control_file_upgrade.rs +++ b/safekeeper/src/control_file_upgrade.rs @@ -1,17 +1,22 @@ //! Code to deal with safekeeper control file upgrades +use std::vec; + use crate::{ safekeeper::{AcceptorState, PgUuid, TermHistory, TermLsn}, - state::{EvictionState, PersistedPeers, TimelinePersistentState}, + state::{EvictionState, TimelinePersistentState}, wal_backup_partial, }; use anyhow::{bail, Result}; use pq_proto::SystemId; -use safekeeper_api::{ServerInfo, Term}; +use safekeeper_api::{ + membership::{Configuration, INVALID_GENERATION}, + ServerInfo, Term, +}; use serde::{Deserialize, Serialize}; use tracing::*; use utils::{ bin_ser::LeSer, - id::{TenantId, TimelineId}, + id::{NodeId, TenantId, TimelineId}, lsn::Lsn, }; @@ -233,6 +238,90 @@ pub struct SafeKeeperStateV8 { pub partial_backup: wal_backup_partial::State, } +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +pub struct PersistedPeers(pub Vec<(NodeId, PersistedPeerInfo)>); + +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +pub struct PersistedPeerInfo { + /// LSN up to which safekeeper offloaded WAL to s3. + pub backup_lsn: Lsn, + /// Term of the last entry. + pub term: Term, + /// LSN of the last record. + pub flush_lsn: Lsn, + /// Up to which LSN safekeeper regards its WAL as committed. + pub commit_lsn: Lsn, +} + +impl PersistedPeerInfo { + pub fn new() -> Self { + Self { + backup_lsn: Lsn::INVALID, + term: safekeeper_api::INITIAL_TERM, + flush_lsn: Lsn(0), + commit_lsn: Lsn(0), + } + } +} + +// make clippy happy +impl Default for PersistedPeerInfo { + fn default() -> Self { + Self::new() + } +} + +/// Note: SafekeeperStateVn is old name for TimelinePersistentStateVn. +#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] +pub struct TimelinePersistentStateV9 { + #[serde(with = "hex")] + pub tenant_id: TenantId, + #[serde(with = "hex")] + pub timeline_id: TimelineId, + /// persistent acceptor state + pub acceptor_state: AcceptorState, + /// information about server + pub server: ServerInfo, + /// Unique id of the last *elected* proposer we dealt with. Not needed + /// for correctness, exists for monitoring purposes. + #[serde(with = "hex")] + pub proposer_uuid: PgUuid, + /// Since which LSN this timeline generally starts. Safekeeper might have + /// joined later. + pub timeline_start_lsn: Lsn, + /// Since which LSN safekeeper has (had) WAL for this timeline. + /// All WAL segments next to one containing local_start_lsn are + /// filled with data from the beginning. + pub local_start_lsn: Lsn, + /// Part of WAL acknowledged by quorum *and available locally*. Always points + /// to record boundary. + pub commit_lsn: Lsn, + /// LSN that points to the end of the last backed up segment. Useful to + /// persist to avoid finding out offloading progress on boot. + pub backup_lsn: Lsn, + /// Minimal LSN which may be needed for recovery of some safekeeper (end_lsn + /// of last record streamed to everyone). Persisting it helps skipping + /// recovery in walproposer, generally we compute it from peers. In + /// walproposer proto called 'truncate_lsn'. Updates are currently drived + /// only by walproposer. + pub peer_horizon_lsn: Lsn, + /// LSN of the oldest known checkpoint made by pageserver and successfully + /// pushed to s3. We don't remove WAL beyond it. Persisted only for + /// informational purposes, we receive it from pageserver (or broker). + pub remote_consistent_lsn: Lsn, + /// Peers and their state as we remember it. Knowing peers themselves is + /// fundamental; but state is saved here only for informational purposes and + /// obviously can be stale. (Currently not saved at all, but let's provision + /// place to have less file version upgrades). + pub peers: PersistedPeers, + /// Holds names of partial segments uploaded to remote storage. Used to + /// clean up old objects without leaving garbage in remote storage. + pub partial_backup: wal_backup_partial::State, + /// Eviction state of the timeline. If it's Offloaded, we should download + /// WAL files from remote storage to serve the timeline. + pub eviction_state: EvictionState, +} + pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result { // migrate to storing full term history if version == 1 { @@ -248,6 +337,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result Result Result Result Result Result Result Result Result Result Result Result Result SafeKeeperStateV8 { - assert!(state.eviction_state == EvictionState::Present); - SafeKeeperStateV8 { +// Used as a temp hack to make forward compatibility test work. Should be +// removed after PR adding v10 is merged. +pub fn downgrade_v10_to_v9(state: &TimelinePersistentState) -> TimelinePersistentStateV9 { + assert!(state.mconf.generation == INVALID_GENERATION); + TimelinePersistentStateV9 { tenant_id: state.tenant_id, timeline_id: state.timeline_id, acceptor_state: state.acceptor_state.clone(), @@ -426,8 +542,9 @@ pub fn downgrade_v9_to_v8(state: &TimelinePersistentState) -> SafeKeeperStateV8 backup_lsn: state.backup_lsn, peer_horizon_lsn: state.peer_horizon_lsn, remote_consistent_lsn: state.remote_consistent_lsn, - peers: state.peers.clone(), + peers: PersistedPeers(vec![]), partial_backup: state.partial_backup.clone(), + eviction_state: state.eviction_state, } } @@ -437,7 +554,7 @@ mod tests { use utils::{id::NodeId, Hex}; - use crate::safekeeper::PersistedPeerInfo; + use crate::control_file_upgrade::PersistedPeerInfo; use super::*; diff --git a/safekeeper/src/copy_timeline.rs b/safekeeper/src/copy_timeline.rs index 28ef2b1d23..10a761e1f5 100644 --- a/safekeeper/src/copy_timeline.rs +++ b/safekeeper/src/copy_timeline.rs @@ -1,6 +1,7 @@ use anyhow::{bail, Result}; use camino::Utf8PathBuf; use postgres_ffi::{MAX_SEND_SIZE, WAL_SEGMENT_SIZE}; +use safekeeper_api::membership::Configuration; use std::sync::Arc; use tokio::{ fs::OpenOptions, @@ -147,10 +148,10 @@ pub async fn handle_request( let mut new_state = TimelinePersistentState::new( &request.destination_ttid, + Configuration::empty(), state.server.clone(), - vec![], - request.until_lsn, start_lsn, + request.until_lsn, )?; new_state.timeline_start_lsn = start_lsn; new_state.peer_horizon_lsn = request.until_lsn; diff --git a/safekeeper/src/http/routes.rs b/safekeeper/src/http/routes.rs index 6186f4c3ba..3835d39698 100644 --- a/safekeeper/src/http/routes.rs +++ b/safekeeper/src/http/routes.rs @@ -111,14 +111,15 @@ async fn timeline_create_handler(mut request: Request) -> Result NetworkReader<'_, IO> { }; let tli = self .global_timelines - .create(self.ttid, server_info, Lsn::INVALID, Lsn::INVALID) + .create( + self.ttid, + Configuration::empty(), + server_info, + Lsn::INVALID, + Lsn::INVALID, + ) .await .context("create timeline")?; tli.wal_residence_guard().await? diff --git a/safekeeper/src/safekeeper.rs b/safekeeper/src/safekeeper.rs index 6ceaf325b0..06403228e9 100644 --- a/safekeeper/src/safekeeper.rs +++ b/safekeeper/src/safekeeper.rs @@ -7,7 +7,6 @@ use bytes::{Buf, BufMut, Bytes, BytesMut}; use postgres_ffi::{TimeLineID, MAX_SEND_SIZE}; use safekeeper_api::models::HotStandbyFeedback; use safekeeper_api::Term; -use safekeeper_api::INVALID_TERM; use serde::{Deserialize, Serialize}; use std::cmp::max; use std::cmp::min; @@ -193,36 +192,6 @@ impl AcceptorState { } } -#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] -pub struct PersistedPeerInfo { - /// LSN up to which safekeeper offloaded WAL to s3. - pub backup_lsn: Lsn, - /// Term of the last entry. - pub term: Term, - /// LSN of the last record. - pub flush_lsn: Lsn, - /// Up to which LSN safekeeper regards its WAL as committed. - pub commit_lsn: Lsn, -} - -impl PersistedPeerInfo { - pub fn new() -> Self { - Self { - backup_lsn: Lsn::INVALID, - term: INVALID_TERM, - flush_lsn: Lsn(0), - commit_lsn: Lsn(0), - } - } -} - -// make clippy happy -impl Default for PersistedPeerInfo { - fn default() -> Self { - Self::new() - } -} - // protocol messages /// Initial Proposer -> Acceptor message @@ -1010,7 +979,7 @@ where /// Update commit_lsn from peer safekeeper data. pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> { - if (Lsn(sk_info.commit_lsn) != Lsn::INVALID) && (sk_info.last_log_term != INVALID_TERM) { + if Lsn(sk_info.commit_lsn) != Lsn::INVALID { // Note: the check is too restrictive, generally we can update local // commit_lsn if our history matches (is part of) history of advanced // commit_lsn provider. @@ -1025,12 +994,20 @@ where #[cfg(test)] mod tests { use futures::future::BoxFuture; + use postgres_ffi::{XLogSegNo, WAL_SEGMENT_SIZE}; - use safekeeper_api::ServerInfo; + use safekeeper_api::{ + membership::{Configuration, MemberSet, SafekeeperId}, + ServerInfo, + }; use super::*; - use crate::state::{EvictionState, PersistedPeers, TimelinePersistentState}; - use std::{ops::Deref, str::FromStr, time::Instant}; + use crate::state::{EvictionState, TimelinePersistentState}; + use std::{ + ops::Deref, + str::FromStr, + time::{Instant, UNIX_EPOCH}, + }; // fake storage for tests struct InMemoryState { @@ -1313,12 +1290,21 @@ mod tests { #[test] fn test_sk_state_bincode_serde_roundtrip() { - use utils::Hex; let tenant_id = TenantId::from_str("cf0480929707ee75372337efaa5ecf96").unwrap(); let timeline_id = TimelineId::from_str("112ded66422aa5e953e5440fa5427ac4").unwrap(); let state = TimelinePersistentState { tenant_id, timeline_id, + mconf: Configuration { + generation: 42, + members: MemberSet::new(vec![SafekeeperId { + id: NodeId(1), + host: "hehe.org".to_owned(), + pg_port: 5432, + }]) + .expect("duplicate member"), + new_members: None, + }, acceptor_state: AcceptorState { term: 42, term_history: TermHistory(vec![TermLsn { @@ -1342,70 +1328,13 @@ mod tests { backup_lsn: Lsn(1234567300), peer_horizon_lsn: Lsn(9999999), remote_consistent_lsn: Lsn(1234560000), - peers: PersistedPeers(vec![( - NodeId(1), - PersistedPeerInfo { - backup_lsn: Lsn(1234567000), - term: 42, - flush_lsn: Lsn(1234567800 - 8), - commit_lsn: Lsn(1234567600), - }, - )]), partial_backup: crate::wal_backup_partial::State::default(), eviction_state: EvictionState::Present, + creation_ts: UNIX_EPOCH, }; let ser = state.ser().unwrap(); - #[rustfmt::skip] - let expected = [ - // tenant_id as length prefixed hex - 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x63, 0x66, 0x30, 0x34, 0x38, 0x30, 0x39, 0x32, 0x39, 0x37, 0x30, 0x37, 0x65, 0x65, 0x37, 0x35, 0x33, 0x37, 0x32, 0x33, 0x33, 0x37, 0x65, 0x66, 0x61, 0x61, 0x35, 0x65, 0x63, 0x66, 0x39, 0x36, - // timeline_id as length prefixed hex - 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x31, 0x31, 0x32, 0x64, 0x65, 0x64, 0x36, 0x36, 0x34, 0x32, 0x32, 0x61, 0x61, 0x35, 0x65, 0x39, 0x35, 0x33, 0x65, 0x35, 0x34, 0x34, 0x30, 0x66, 0x61, 0x35, 0x34, 0x32, 0x37, 0x61, 0x63, 0x34, - // term - 0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // length prefix - 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // unsure why this order is swapped - 0x29, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // pg_version - 0x0e, 0x00, 0x00, 0x00, - // systemid - 0x21, 0x43, 0x65, 0x87, 0x78, 0x56, 0x34, 0x12, - // wal_seg_size - 0x78, 0x56, 0x34, 0x12, - // pguuid as length prefixed hex - 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x63, 0x34, 0x37, 0x61, 0x34, 0x32, 0x61, 0x35, 0x30, 0x66, 0x34, 0x34, 0x65, 0x35, 0x35, 0x33, 0x65, 0x39, 0x61, 0x35, 0x32, 0x61, 0x34, 0x32, 0x36, 0x36, 0x65, 0x64, 0x32, 0x64, 0x31, 0x31, - - // timeline_start_lsn - 0x00, 0x56, 0x34, 0x12, 0x00, 0x00, 0x00, 0x00, - 0x12, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x78, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00, - 0x84, 0x00, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00, - 0x7f, 0x96, 0x98, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x00, 0xe4, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00, - // length prefix for persistentpeers - 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // nodeid - 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // backuplsn - 0x58, 0xff, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00, - 0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - 0x70, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00, - 0xb0, 0x01, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00, - // partial_backup - 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, - // eviction_state - 0x00, 0x00, 0x00, 0x00, - ]; - - assert_eq!(Hex(&ser), Hex(&expected)); - let deser = TimelinePersistentState::des(&ser).unwrap(); assert_eq!(deser, state); diff --git a/safekeeper/src/state.rs b/safekeeper/src/state.rs index c6ae6c1d2b..1c3bb1b4dc 100644 --- a/safekeeper/src/state.rs +++ b/safekeeper/src/state.rs @@ -1,20 +1,22 @@ //! Defines per timeline data stored persistently (SafeKeeperPersistentState) //! and its wrapper with in memory layer (SafekeeperState). -use std::{cmp::max, ops::Deref}; +use std::{cmp::max, ops::Deref, time::SystemTime}; use anyhow::{bail, Result}; use postgres_ffi::WAL_SEGMENT_SIZE; -use safekeeper_api::{models::TimelineTermBumpResponse, ServerInfo, Term}; +use safekeeper_api::{ + membership::Configuration, models::TimelineTermBumpResponse, ServerInfo, Term, INITIAL_TERM, +}; use serde::{Deserialize, Serialize}; use utils::{ - id::{NodeId, TenantId, TenantTimelineId, TimelineId}, + id::{TenantId, TenantTimelineId, TimelineId}, lsn::Lsn, }; use crate::{ control_file, - safekeeper::{AcceptorState, PersistedPeerInfo, PgUuid, TermHistory, UNKNOWN_SERVER_VERSION}, + safekeeper::{AcceptorState, PgUuid, TermHistory, TermLsn, UNKNOWN_SERVER_VERSION}, timeline::TimelineError, wal_backup_partial::{self}, }; @@ -27,6 +29,8 @@ pub struct TimelinePersistentState { pub tenant_id: TenantId, #[serde(with = "hex")] pub timeline_id: TimelineId, + /// Membership configuration. + pub mconf: Configuration, /// persistent acceptor state pub acceptor_state: AcceptorState, /// information about server @@ -58,22 +62,15 @@ pub struct TimelinePersistentState { /// pushed to s3. We don't remove WAL beyond it. Persisted only for /// informational purposes, we receive it from pageserver (or broker). pub remote_consistent_lsn: Lsn, - /// Peers and their state as we remember it. Knowing peers themselves is - /// fundamental; but state is saved here only for informational purposes and - /// obviously can be stale. (Currently not saved at all, but let's provision - /// place to have less file version upgrades). - pub peers: PersistedPeers, /// Holds names of partial segments uploaded to remote storage. Used to /// clean up old objects without leaving garbage in remote storage. pub partial_backup: wal_backup_partial::State, /// Eviction state of the timeline. If it's Offloaded, we should download /// WAL files from remote storage to serve the timeline. pub eviction_state: EvictionState, + pub creation_ts: SystemTime, } -#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)] -pub struct PersistedPeers(pub Vec<(NodeId, PersistedPeerInfo)>); - /// State of the local WAL files. Used to track current timeline state, /// that can be either WAL files are present on disk or last partial segment /// is offloaded to remote storage. @@ -87,12 +84,14 @@ pub enum EvictionState { } impl TimelinePersistentState { + /// commit_lsn is the same as start_lsn in the normal creaiton; see + /// `TimelineCreateRequest` comments.` pub fn new( ttid: &TenantTimelineId, + mconf: Configuration, server_info: ServerInfo, - peers: Vec, + start_lsn: Lsn, commit_lsn: Lsn, - local_start_lsn: Lsn, ) -> anyhow::Result { if server_info.wal_seg_size == 0 { bail!(TimelineError::UninitializedWalSegSize(*ttid)); @@ -102,49 +101,59 @@ impl TimelinePersistentState { bail!(TimelineError::UninitialinzedPgVersion(*ttid)); } - if commit_lsn < local_start_lsn { + if commit_lsn < start_lsn { bail!( - "commit_lsn {} is smaller than local_start_lsn {}", + "commit_lsn {} is smaller than start_lsn {}", commit_lsn, - local_start_lsn + start_lsn ); } + // If we are given with init LSN, initialize term history with it. It + // ensures that walproposer always must be able to find a common point + // in histories; if it can't something is corrupted. Not having LSN here + // is so far left for legacy case where timeline is created by compute + // and LSN during creation is not known yet. + let term_history = if commit_lsn != Lsn::INVALID { + TermHistory(vec![TermLsn { + term: INITIAL_TERM, + lsn: start_lsn, + }]) + } else { + TermHistory::empty() + }; + Ok(TimelinePersistentState { tenant_id: ttid.tenant_id, timeline_id: ttid.timeline_id, + mconf, acceptor_state: AcceptorState { - term: 0, - term_history: TermHistory::empty(), + term: INITIAL_TERM, + term_history, }, server: server_info, proposer_uuid: [0; 16], - timeline_start_lsn: Lsn(0), - local_start_lsn, + timeline_start_lsn: start_lsn, + local_start_lsn: start_lsn, commit_lsn, - backup_lsn: local_start_lsn, - peer_horizon_lsn: local_start_lsn, + backup_lsn: start_lsn, + peer_horizon_lsn: start_lsn, remote_consistent_lsn: Lsn(0), - peers: PersistedPeers( - peers - .iter() - .map(|p| (*p, PersistedPeerInfo::new())) - .collect(), - ), partial_backup: wal_backup_partial::State::default(), eviction_state: EvictionState::Present, + creation_ts: SystemTime::now(), }) } pub fn empty() -> Self { TimelinePersistentState::new( &TenantTimelineId::empty(), + Configuration::empty(), ServerInfo { pg_version: 170000, /* Postgres server version (major * 10000) */ system_id: 0, /* Postgres system identifier */ wal_seg_size: WAL_SEGMENT_SIZE as u32, }, - vec![], Lsn::INVALID, Lsn::INVALID, ) diff --git a/safekeeper/src/timelines_global_map.rs b/safekeeper/src/timelines_global_map.rs index ad29c9f66c..a701534f65 100644 --- a/safekeeper/src/timelines_global_map.rs +++ b/safekeeper/src/timelines_global_map.rs @@ -12,6 +12,7 @@ use crate::{control_file, wal_storage, SafeKeeperConf}; use anyhow::{bail, Context, Result}; use camino::Utf8PathBuf; use camino_tempfile::Utf8TempDir; +use safekeeper_api::membership::Configuration; use safekeeper_api::ServerInfo; use serde::Serialize; use std::collections::HashMap; @@ -214,9 +215,10 @@ impl GlobalTimelines { pub(crate) async fn create( &self, ttid: TenantTimelineId, + mconf: Configuration, server_info: ServerInfo, + start_lsn: Lsn, commit_lsn: Lsn, - local_start_lsn: Lsn, ) -> Result> { let (conf, _, _) = { let state = self.state.lock().unwrap(); @@ -239,8 +241,7 @@ impl GlobalTimelines { // TODO: currently we create only cfile. It would be reasonable to // immediately initialize first WAL segment as well. - let state = - TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?; + let state = TimelinePersistentState::new(&ttid, mconf, server_info, start_lsn, commit_lsn)?; control_file::FileStorage::create_new(&tmp_dir_path, state, conf.no_sync).await?; let timeline = self.load_temp_timeline(ttid, &tmp_dir_path, true).await?; Ok(timeline) diff --git a/safekeeper/tests/walproposer_sim/safekeeper.rs b/safekeeper/tests/walproposer_sim/safekeeper.rs index efcdd89e7d..a99de71a04 100644 --- a/safekeeper/tests/walproposer_sim/safekeeper.rs +++ b/safekeeper/tests/walproposer_sim/safekeeper.rs @@ -21,7 +21,7 @@ use safekeeper::{ wal_storage::Storage, SafeKeeperConf, }; -use safekeeper_api::ServerInfo; +use safekeeper_api::{membership::Configuration, ServerInfo}; use tracing::{debug, info_span, warn}; use utils::{ id::{NodeId, TenantId, TenantTimelineId, TimelineId}, @@ -96,8 +96,13 @@ impl GlobalMap { let commit_lsn = Lsn::INVALID; let local_start_lsn = Lsn::INVALID; - let state = - TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?; + let state = TimelinePersistentState::new( + &ttid, + Configuration::empty(), + server_info, + commit_lsn, + local_start_lsn, + )?; let disk_timeline = self.disk.put_state(&ttid, state); let control_store = DiskStateStorage::new(disk_timeline.clone()); diff --git a/test_runner/fixtures/pageserver/http.py b/test_runner/fixtures/pageserver/http.py index 378e568622..364aff325d 100644 --- a/test_runner/fixtures/pageserver/http.py +++ b/test_runner/fixtures/pageserver/http.py @@ -15,7 +15,6 @@ from requests.adapters import HTTPAdapter from urllib3.util.retry import Retry from fixtures.common_types import ( - Id, Lsn, TenantId, TenantShardId, @@ -25,7 +24,7 @@ from fixtures.common_types import ( from fixtures.log_helper import log from fixtures.metrics import Metrics, MetricsGetter, parse_metrics from fixtures.pg_version import PgVersion -from fixtures.utils import Fn +from fixtures.utils import EnhancedJSONEncoder, Fn class PageserverApiException(Exception): @@ -83,14 +82,6 @@ class TimelineCreateRequest: mode: TimelineCreateRequestMode def to_json(self) -> str: - class EnhancedJSONEncoder(json.JSONEncoder): - def default(self, o): - if dataclasses.is_dataclass(o) and not isinstance(o, type): - return dataclasses.asdict(o) - elif isinstance(o, Id): - return o.id.hex() - return super().default(o) - # mode is flattened this = dataclasses.asdict(self) mode = this.pop("mode") diff --git a/test_runner/fixtures/safekeeper/http.py b/test_runner/fixtures/safekeeper/http.py index 286f80ba69..4826cae3ee 100644 --- a/test_runner/fixtures/safekeeper/http.py +++ b/test_runner/fixtures/safekeeper/http.py @@ -10,7 +10,7 @@ import requests from fixtures.common_types import Lsn, TenantId, TenantTimelineId, TimelineId from fixtures.log_helper import log from fixtures.metrics import Metrics, MetricsGetter, parse_metrics -from fixtures.utils import wait_until +from fixtures.utils import EnhancedJSONEncoder, wait_until if TYPE_CHECKING: from typing import Any @@ -69,6 +69,34 @@ class TermBumpResponse: ) +@dataclass +class SafekeeperId: + id: int + host: str + pg_port: str + + +@dataclass +class Configuration: + generation: int + members: list[SafekeeperId] + new_members: list[SafekeeperId] | None + + +@dataclass +class TimelineCreateRequest: + tenant_id: TenantId + timeline_id: TimelineId + mconf: Configuration + # not exactly PgVersion, for example 150002 for 15.2 + pg_version: int + start_lsn: Lsn + commit_lsn: Lsn | None + + def to_json(self) -> str: + return json.dumps(self, cls=EnhancedJSONEncoder) + + class SafekeeperHttpClient(requests.Session, MetricsGetter): HTTPError = requests.HTTPError @@ -131,20 +159,8 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter): resj = res.json() return [TenantTimelineId.from_json(ttidj) for ttidj in resj] - def timeline_create( - self, - tenant_id: TenantId, - timeline_id: TimelineId, - pg_version: int, # Not exactly a PgVersion, safekeeper returns version as int, for example 150002 for 15.2 - commit_lsn: Lsn, - ): - body = { - "tenant_id": str(tenant_id), - "timeline_id": str(timeline_id), - "pg_version": pg_version, - "commit_lsn": str(commit_lsn), - } - res = self.post(f"http://localhost:{self.port}/v1/tenant/timeline", json=body) + def timeline_create(self, r: TimelineCreateRequest): + res = self.post(f"http://localhost:{self.port}/v1/tenant/timeline", data=r.to_json()) res.raise_for_status() def timeline_status( diff --git a/test_runner/fixtures/utils.py b/test_runner/fixtures/utils.py index c34ac298d1..e160c617cd 100644 --- a/test_runner/fixtures/utils.py +++ b/test_runner/fixtures/utils.py @@ -1,6 +1,7 @@ from __future__ import annotations import contextlib +import dataclasses import json import os import re @@ -21,6 +22,7 @@ import zstandard from psycopg2.extensions import cursor from typing_extensions import override +from fixtures.common_types import Id, Lsn from fixtures.log_helper import log from fixtures.pageserver.common_types import ( parse_delta_layer, @@ -605,6 +607,22 @@ class PropagatingThread(threading.Thread): return self.ret +class EnhancedJSONEncoder(json.JSONEncoder): + """ + Default json.JSONEncoder works only on primitive builtins. Extend it to any + dataclass plus our custom types. + """ + + def default(self, o): + if dataclasses.is_dataclass(o) and not isinstance(o, type): + return dataclasses.asdict(o) + elif isinstance(o, Id): + return o.id.hex() + elif isinstance(o, Lsn): + return str(o) # standard hex notation + return super().default(o) + + def human_bytes(amt: float) -> str: """ Render a bytes amount into nice IEC bytes string. diff --git a/test_runner/regress/test_wal_acceptor.py b/test_runner/regress/test_wal_acceptor.py index 0a8900b351..d39c6a6b5b 100644 --- a/test_runner/regress/test_wal_acceptor.py +++ b/test_runner/regress/test_wal_acceptor.py @@ -48,7 +48,7 @@ from fixtures.remote_storage import ( default_remote_storage, s3_storage, ) -from fixtures.safekeeper.http import SafekeeperHttpClient +from fixtures.safekeeper.http import Configuration, SafekeeperHttpClient, TimelineCreateRequest from fixtures.safekeeper.utils import wait_walreceivers_absent from fixtures.utils import ( PropagatingThread, @@ -658,7 +658,13 @@ def test_s3_wal_replay(neon_env_builder: NeonEnvBuilder): for sk in env.safekeepers: sk.start() cli = sk.http_client() - cli.timeline_create(tenant_id, timeline_id, pg_version, last_lsn) + mconf = Configuration(generation=0, members=[], new_members=None) + # set start_lsn to the beginning of the first segment to allow reading + # WAL from there (could you intidb LSN as well). + r = TimelineCreateRequest( + tenant_id, timeline_id, mconf, pg_version, Lsn("0/1000000"), commit_lsn=last_lsn + ) + cli.timeline_create(r) f_partial_path = ( Path(sk.data_dir) / str(tenant_id) / str(timeline_id) / f_partial_saved.name ) From 05d17a10aee949b56317bc1ecaf24b76b097b9d2 Mon Sep 17 00:00:00 2001 From: Erik Grinaker Date: Wed, 15 Jan 2025 11:35:38 +0100 Subject: [PATCH 12/31] rfc: add CPU and heap profiling RFC (#10085) This document proposes a standard cross-team pattern for CPU and memory profiling across applications and languages, using the [pprof](https://github.com/google/pprof) profile format. It enables both ad hoc profiles via HTTP endpoints, and continuous profiling across the fleet via [Grafana Cloud Profiles](https://grafana.com/docs/grafana-cloud/monitor-applications/profiles/). Continuous profiling incurs an overhead of about 0.1% CPU usage and 3% slower heap allocations. [Rendered](https://github.com/neondatabase/neon/blob/erik/profiling-rfc/docs/rfcs/040-profiling.md) Touches #9534. Touches https://github.com/neondatabase/cloud/issues/14888. --- docs/rfcs/040-profiling.md | 247 +++++++++++++++++++++++++++++++++++++ 1 file changed, 247 insertions(+) create mode 100644 docs/rfcs/040-profiling.md diff --git a/docs/rfcs/040-profiling.md b/docs/rfcs/040-profiling.md new file mode 100644 index 0000000000..8da9e50774 --- /dev/null +++ b/docs/rfcs/040-profiling.md @@ -0,0 +1,247 @@ +# CPU and Memory Profiling + +Created 2025-01-12 by Erik Grinaker. + +See also [internal user guide](https://www.notion.so/neondatabase/Storage-CPU-Memory-Profiling-14bf189e004780228ec7d04442742324?pvs=4). + +## Summary + +This document proposes a standard cross-team pattern for CPU and memory profiling across +applications and languages, using the [pprof](https://github.com/google/pprof) profile format. + +It enables both ad hoc profiles via HTTP endpoints, and continuous profiling across the fleet via +[Grafana Cloud Profiles](https://grafana.com/docs/grafana-cloud/monitor-applications/profiles/). +Continuous profiling incurs an overhead of about 0.1% CPU usage and 3% slower heap allocations. + +## Motivation + +CPU and memory profiles are crucial observability tools for understanding performance issues, +resource exhaustion, and resource costs. They allow answering questions like: + +* Why is this process using 100% CPU? +* How do I make this go faster? +* Why did this process run out of memory? +* Why are we paying for all these CPU cores and memory chips? + +Go has [first-class support](https://pkg.go.dev/net/http/pprof) for profiling included in its +standard library, using the [pprof](https://github.com/google/pprof) profile format and associated +tooling. + +This is not the case for Rust and C, where obtaining profiles can be rather cumbersome. It requires +installing and running additional tools like `perf` as root on production nodes, with analysis tools +that can be hard to use and often don't give good results. This is not only annoying, but can also +significantly affect the resolution time of production incidents. + +This proposal will: + +* Provide CPU and heap profiles in pprof format via HTTP API. +* Record continuous profiles in Grafana for aggregate historical analysis. +* Make it easy for anyone to see a flamegraph in less than one minute. +* Be reasonably consistent across teams and services (Rust, Go, C). + +## Non Goals (For Now) + +* [Additional profile types](https://grafana.com/docs/pyroscope/next/configure-client/profile-types/) + like mutexes, locks, goroutines, etc. +* [Runtime trace integration](https://grafana.com/docs/pyroscope/next/configure-client/trace-span-profiles/). +* [Profile-guided optimization](https://en.wikipedia.org/wiki/Profile-guided_optimization). + +## Using Profiles + +Ready-to-use profiles can be obtained using e.g. `curl`. For Rust services: + +``` +$ curl localhost:9898/profile/cpu >profile.pb.gz +``` + +pprof profiles can be explored using the [`pprof`](https://github.com/google/pprof) web UI, which +provides flamegraphs, call graphs, plain text listings, and more: + +``` +$ pprof -http :6060 +``` + +Some endpoints (e.g. Rust-based ones) can also generate flamegraph SVGs directly: + +``` +$ curl localhost:9898/profile/cpu?format=svg >profile.svg +$ open profile.svg +``` + +Continuous profiles are available in Grafana under Explore → Profiles → Explore Profiles +(currently only in [staging](https://neonstaging.grafana.net/a/grafana-pyroscope-app/profiles-explorer)). + +## API Requirements + +* HTTP endpoints that return a profile in pprof format (with symbols). + * CPU: records a profile over the request time interval (`seconds` query parameter). + * Memory: returns the current in-use heap allocations. +* Unauthenticated, as it should not expose user data or pose a denial-of-service risk. +* Default sample frequency should not impact service (maximum 5% CPU overhead). +* Linux-compatibility. + +Nice to have: + +* Return flamegraph SVG directly from the HTTP endpoint if requested. +* Configurable sample frequency for CPU profiles. +* Historical heap allocations, by count and bytes. +* macOS-compatiblity. + +## Rust Profiling + +[`libs/utils/src/http/endpoint.rs`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs) +contains ready-to-use HTTP endpoints for CPU and memory profiling: +[`profile_cpu_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L338) and [`profile_heap_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L416). + +### CPU + +CPU profiles are provided by [pprof-rs](https://github.com/tikv/pprof-rs) via +[`profile_cpu_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L338). +Expose it unauthenticated at `/profile/cpu`. + +Parameters: + +* `format`: profile output format (`pprof` or `svg`; default `pprof`). +* `seconds`: duration to collect profile over, in seconds (default `5`). +* `frequency`: how often to sample thread stacks, in Hz (default `99`). +* `force`: if `true`, cancel a running profile and start a new one (default `false`). + +Works on Linux and macOS. + +### Memory + +Use the jemalloc allocator via [`tikv-jemallocator`](https://github.com/tikv/jemallocator), +and enable profiling with samples every 2 MB allocated: + +```rust +#[global_allocator] +static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc; + +#[allow(non_upper_case_globals)] +#[export_name = "malloc_conf"] +pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0"; +``` + +pprof profiles are generated by +[`jemalloc-pprof`](https://github.com/polarsignals/rust-jemalloc-pprof) via +[`profile_heap_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L416). +Expose it unauthenticated at `/profile/heap`. + +Parameters: + +* `format`: profile output format (`pprof`, `svg`, or `jemalloc`; default `pprof`). + +Works on Linux only, due to [jemalloc limitations](https://github.com/jemalloc/jemalloc/issues/26). + +## Go Profiling + +The Go standard library includes pprof profiling via HTTP API in +[`net/http/pprof`](https://pkg.go.dev/net/http/pprof). Expose it unauthenticated at +`/debug/pprof`. + +Works on Linux and macOS. + +### CPU + +Via `/debug/pprof/profile`. Parameters: + +* `debug`: profile output format (`0` is pprof, `1` or above is plaintext; default `0`). +* `seconds`: duration to collect profile over, in seconds (default `30`). + +Does not support a frequency parameter (see [#57488](https://github.com/golang/go/issues/57488)), +and defaults to 100 Hz. A lower frequency can be hardcoded via `SetCPUProfileRate`, but the default +is likely ok (estimated 1% overhead). + +### Memory + +Via `/debug/pprof/heap`. Parameters: + +* `seconds`: take a delta profile over the given duration, in seconds (default `0`). +* `gc`: if `1`, garbage collect before taking profile. + +## C Profiling + +[gperftools](https://github.com/gperftools/gperftools) provides in-process CPU and heap profiling +with pprof output. + +However, continuous profiling of PostgreSQL is expensive (many computes), and has limited value +since we don't own the internals anyway. + +Ad hoc profiling might still be useful, but the compute team considers existing tooling sufficient, +so this is not a priority at the moment. + +## Grafana Continuous Profiling + +[Grafana Alloy](https://grafana.com/docs/alloy/latest/) continually scrapes CPU and memory profiles +across the fleet, and archives them as time series. This can be used to analyze resource usage over +time, either in aggregate or zoomed in to specific events and nodes. + +Profiles are retained for 30 days. Profile ingestion volume for CPU+heap at 60-second intervals +is about 0.5 GB/node/day, or about $0.25/node/day = $7.5/node/month ($0.50/GB). + +It is currently enabled in [staging](https://neonstaging.grafana.net/a/grafana-pyroscope-app/profiles-explorer) +for Pageserver and Safekeeper. + +### Scraping + +* CPU profiling: 59 seconds at 19 Hz every 60 seconds. +* Heap profiling: heap snapshot with 2 MB frequency every 60 seconds. + +There are two main approaches that can be taken for CPU profiles: + +* Continuous low-frequency profiles (e.g. 19 Hz for 60 seconds every 60 seconds). +* Occasional high-frequency profiles (e.g. 99 Hz for 5 seconds every 60 seconds). + +We choose continuous low-frequency profiles where possible. This has a fixed low overhead, instead +of a spiky high overhead. It likely also gives a more representative view of resource usage. +However, a 19 Hz rate gives a minimum resolution of 52.6 ms per sample, which may be larger than the +actual runtime of small functions. Note that Go does not support a frequency parameter, so we must +use a fixed frequency for all profiles via `SetCPUProfileRate()` (default 100 Hz). + +Only one CPU profile can be taken at a time. With continuous profiling, one will always be running. +To allow also taking an ad hoc CPU profile, the Rust endpoint supports a `force` query parameter to +cancel a running profile and start a new one. + +### Overhead + +With Rust: + +* CPU profiles at 19 Hz frequency: 0.1% overhead. +* Heap profiles at 2 MB frequency: 3% allocation overhead. +* Profile call/encoding/symbolization: 20 ms every 60 seconds, or 0.03% of 1 CPU (for Pageserver). +* Profile symbolization caches: 125 MB memory, or 0.4% of 32 GB (for Pageserver). + +Benchmarks with pprof-rs showed that the CPU time for taking a stack trace of a 40-frame stack was +11 µs using the `frame-pointer` feature, and 1.4 µs using `libunwind` with DWARF. `libunwind` saw +frequent seg faults, so we use `frame-pointer` and build binaries with frame pointers (negligible +overhead). + +CPU profiles work by installing an `ITIMER_PROF` for the process, which triggers a `SIGPROF` signal +after a given amount of cumulative CPU time across all CPUs. The signal handler will run for one +of the currently executing threads and take a stack trace. Thus, a 19 Hz profile will take 1 stack +trace every 52.6 ms CPU time -- assuming 11 µs for a stack trace, this is 0.02% overhead, but +likely 0.1% in practice (given e.g. context switches). + +Heap profiles work by probabilistically taking a stack trace on allocations, adjusted for the +allocation size. A 1 MB allocation takes about 15 µs in benchmarks, and a stack trace about 1 µs, +so we can estimate that a 2 MB sampling frequency has about 3% allocation overhead -- this is +consistent with benchmarks. This is significantly larger than CPU profiles, but mitigated by the +fact that performance-sensitive code will avoid allocations as far as possible. + +Profile symbolization uses in-memory caches for symbol lookups. These take about 125 MB for +Pageserver. + +## Alternatives Considered + +* eBPF profiles. + * Don't require instrumenting the binary. + * Use less resources. + * Can profile in kernel space too. + * Supported by Grafana. + * Less information about stack frames and spans. + * Limited tooling for local analysis. + * Does not support heap profiles. + * Does not work on macOS. + +* [Polar Signals](https://www.polarsignals.com) instead of Grafana. + * We already use Grafana for everything else. Appears good enough. From 157743040896e59214260a78542899c07638f94d Mon Sep 17 00:00:00 2001 From: Vlad Lazar Date: Wed, 15 Jan 2025 11:10:24 +0000 Subject: [PATCH 13/31] safekeeper: decode and interpret for multiple shards in one go (#10201) ## Problem Currently, we call `InterpretedWalRecord::from_bytes_filtered` from each shard. To serve multiple shards at the same time, the API needs to allow for enquiring about multiple shards. ## Summary of changes This commit tweaks it a pretty brute force way. Naively, we could just generate the shard for a key, but pre and post split shards may be subscribed at the same time, so doing it efficiently is more complex. --- Cargo.lock | 9 + libs/pageserver_api/src/shard.rs | 21 +- libs/postgres_ffi/src/walrecord.rs | 14 +- libs/wal_decoder/Cargo.toml | 15 ++ libs/wal_decoder/benches/README.md | 34 +++ .../benches/bench_interpret_wal.rs | 250 ++++++++++++++++++ libs/wal_decoder/src/decoder.rs | 140 ++++++---- libs/wal_decoder/src/models.rs | 54 ++-- libs/wal_decoder/src/serialized_batch.rs | 191 ++++++------- pageserver/src/import_datadir.rs | 15 +- .../walreceiver/walreceiver_connection.rs | 8 +- pageserver/src/walingest.rs | 4 +- safekeeper/src/send_interpreted_wal.rs | 8 +- 13 files changed, 574 insertions(+), 189 deletions(-) create mode 100644 libs/wal_decoder/benches/README.md create mode 100644 libs/wal_decoder/benches/bench_interpret_wal.rs diff --git a/Cargo.lock b/Cargo.lock index f543057933..0669899617 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -7560,12 +7560,21 @@ dependencies = [ "anyhow", "async-compression", "bytes", + "camino", + "camino-tempfile", + "criterion", + "futures", "pageserver_api", "postgres_ffi", + "pprof", "prost", + "remote_storage", "serde", + "serde_json", "thiserror", + "tikv-jemallocator", "tokio", + "tokio-util", "tonic", "tonic-build", "tracing", diff --git a/libs/pageserver_api/src/shard.rs b/libs/pageserver_api/src/shard.rs index 4cc0a739e8..e03df02afb 100644 --- a/libs/pageserver_api/src/shard.rs +++ b/libs/pageserver_api/src/shard.rs @@ -31,6 +31,8 @@ //! - In a tenant with 4 shards, each shard has ShardCount(N), ShardNumber(i) where i in 0..N-1 (inclusive), //! and their slugs are 0004, 0104, 0204, and 0304. +use std::hash::{Hash, Hasher}; + use crate::{key::Key, models::ShardParameters}; use postgres_ffi::relfile_utils::INIT_FORKNUM; use serde::{Deserialize, Serialize}; @@ -48,6 +50,23 @@ pub struct ShardIdentity { layout: ShardLayout, } +/// Hash implementation +/// +/// The stripe size cannot change dynamically, so it can be ignored for efficiency reasons. +impl Hash for ShardIdentity { + fn hash(&self, state: &mut H) { + let ShardIdentity { + number, + count, + stripe_size: _, + layout: _, + } = self; + + number.0.hash(state); + count.0.hash(state); + } +} + /// Stripe size in number of pages #[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)] pub struct ShardStripeSize(pub u32); @@ -59,7 +78,7 @@ impl Default for ShardStripeSize { } /// Layout version: for future upgrades where we might change how the key->shard mapping works -#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)] +#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Hash, Debug)] pub struct ShardLayout(u8); const LAYOUT_V1: ShardLayout = ShardLayout(1); diff --git a/libs/postgres_ffi/src/walrecord.rs b/libs/postgres_ffi/src/walrecord.rs index b32106632a..fce37e2fdd 100644 --- a/libs/postgres_ffi/src/walrecord.rs +++ b/libs/postgres_ffi/src/walrecord.rs @@ -16,7 +16,7 @@ use utils::bin_ser::DeserializeError; use utils::lsn::Lsn; #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlMultiXactCreate { pub mid: MultiXactId, /* new MultiXact's ID */ @@ -46,7 +46,7 @@ impl XlMultiXactCreate { } #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlMultiXactTruncate { pub oldest_multi_db: Oid, /* to-be-truncated range of multixact offsets */ @@ -72,7 +72,7 @@ impl XlMultiXactTruncate { } #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlRelmapUpdate { pub dbid: Oid, /* database ID, or 0 for shared map */ pub tsid: Oid, /* database's tablespace, or pg_global */ @@ -90,7 +90,7 @@ impl XlRelmapUpdate { } #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlReploriginDrop { pub node_id: RepOriginId, } @@ -104,7 +104,7 @@ impl XlReploriginDrop { } #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlReploriginSet { pub remote_lsn: Lsn, pub node_id: RepOriginId, @@ -911,7 +911,7 @@ impl XlSmgrCreate { } #[repr(C)] -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlSmgrTruncate { pub blkno: BlockNumber, pub rnode: RelFileNode, @@ -984,7 +984,7 @@ impl XlDropDatabase { /// xl_xact_parsed_abort structs in PostgreSQL, but we use the same /// struct for commits and aborts. /// -#[derive(Debug, Serialize, Deserialize)] +#[derive(Clone, Debug, Serialize, Deserialize)] pub struct XlXactParsedRecord { pub xid: TransactionId, pub info: u8, diff --git a/libs/wal_decoder/Cargo.toml b/libs/wal_decoder/Cargo.toml index 8fac4e38ca..09c4afb18a 100644 --- a/libs/wal_decoder/Cargo.toml +++ b/libs/wal_decoder/Cargo.toml @@ -24,3 +24,18 @@ workspace_hack = { version = "0.1", path = "../../workspace_hack" } [build-dependencies] tonic-build.workspace = true + +[dev-dependencies] +criterion.workspace = true +camino.workspace = true +camino-tempfile.workspace = true +remote_storage.workspace = true +tokio-util.workspace = true +serde_json.workspace = true +futures.workspace = true +tikv-jemallocator.workspace = true +pprof.workspace = true + +[[bench]] +name = "bench_interpret_wal" +harness = false diff --git a/libs/wal_decoder/benches/README.md b/libs/wal_decoder/benches/README.md new file mode 100644 index 0000000000..14885afecf --- /dev/null +++ b/libs/wal_decoder/benches/README.md @@ -0,0 +1,34 @@ +## WAL Decoding and Interpretation Benchmarks + +Note that these benchmarks pull WAL from a public bucket in S3 +as a preparation step. Hence, you need a way to auth with AWS. +You can achieve this by copying the `~/.aws/config` file from +the AWS SSO notion page and exporting `AWS_PROFILE=dev` when invoking +the benchmarks. + +To run benchmarks: + +```sh +aws sso login --profile dev + +# All benchmarks. +AWS_PROFILE=dev cargo bench --package wal_decoder + +# Specific file. +AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal + +# Specific benchmark. +AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal unsharded + +# List available benchmarks. +cargo bench --package wal_decoder --benches -- --list + +# Generate flamegraph profiles using pprof-rs, profiling for 10 seconds. +# Output in target/criterion/*/profile/flamegraph.svg. +AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal unsharded -- --profile-time 10 +``` + +Additional charts and statistics are available in `target/criterion/report/index.html`. + +Benchmarks are automatically compared against the previous run. To compare against other runs, see +`--baseline` and `--save-baseline`. diff --git a/libs/wal_decoder/benches/bench_interpret_wal.rs b/libs/wal_decoder/benches/bench_interpret_wal.rs new file mode 100644 index 0000000000..846904cf87 --- /dev/null +++ b/libs/wal_decoder/benches/bench_interpret_wal.rs @@ -0,0 +1,250 @@ +use anyhow::Context; +use criterion::{criterion_group, criterion_main, Criterion}; +use futures::{stream::FuturesUnordered, StreamExt}; +use pageserver_api::shard::{ShardIdentity, ShardStripeSize}; +use postgres_ffi::{waldecoder::WalStreamDecoder, MAX_SEND_SIZE, WAL_SEGMENT_SIZE}; +use pprof::criterion::{Output, PProfProfiler}; +use serde::Deserialize; +use std::{env, num::NonZeroUsize, sync::Arc}; + +use camino::{Utf8Path, Utf8PathBuf}; +use camino_tempfile::Utf8TempDir; +use remote_storage::{ + DownloadOpts, GenericRemoteStorage, ListingMode, RemoteStorageConfig, RemoteStorageKind, + S3Config, +}; +use tokio_util::sync::CancellationToken; +use utils::{ + lsn::Lsn, + shard::{ShardCount, ShardNumber}, +}; +use wal_decoder::models::InterpretedWalRecord; + +const S3_BUCKET: &str = "neon-github-public-dev"; +const S3_REGION: &str = "eu-central-1"; +const BUCKET_PREFIX: &str = "wal-snapshots/bulk-insert/"; +const METADATA_FILENAME: &str = "metadata.json"; + +/// Use jemalloc, and configure it to sample allocations for profiles every 1 MB. +/// This mirrors the configuration in bin/safekeeper.rs. +#[global_allocator] +static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc; + +#[allow(non_upper_case_globals)] +#[export_name = "malloc_conf"] +pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:20\0"; + +async fn create_s3_client() -> anyhow::Result> { + let remote_storage_config = RemoteStorageConfig { + storage: RemoteStorageKind::AwsS3(S3Config { + bucket_name: S3_BUCKET.to_string(), + bucket_region: S3_REGION.to_string(), + prefix_in_bucket: Some(BUCKET_PREFIX.to_string()), + endpoint: None, + concurrency_limit: NonZeroUsize::new(100).unwrap(), + max_keys_per_list_response: None, + upload_storage_class: None, + }), + timeout: RemoteStorageConfig::DEFAULT_TIMEOUT, + small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT, + }; + Ok(Arc::new( + GenericRemoteStorage::from_config(&remote_storage_config) + .await + .context("remote storage init")?, + )) +} + +async fn download_bench_data( + client: Arc, + cancel: &CancellationToken, +) -> anyhow::Result { + let temp_dir_parent: Utf8PathBuf = env::current_dir().unwrap().try_into()?; + let temp_dir = camino_tempfile::tempdir_in(temp_dir_parent)?; + + eprintln!("Downloading benchmark data to {:?}", temp_dir); + + let listing = client + .list(None, ListingMode::NoDelimiter, None, cancel) + .await?; + + let mut downloads = listing + .keys + .into_iter() + .map(|obj| { + let client = client.clone(); + let temp_dir_path = temp_dir.path().to_owned(); + + async move { + let remote_path = obj.key; + let download = client + .download(&remote_path, &DownloadOpts::default(), cancel) + .await?; + let mut body = tokio_util::io::StreamReader::new(download.download_stream); + + let file_name = remote_path.object_name().unwrap(); + let file_path = temp_dir_path.join(file_name); + let file = tokio::fs::OpenOptions::new() + .create(true) + .truncate(true) + .write(true) + .open(&file_path) + .await?; + + let mut writer = tokio::io::BufWriter::new(file); + tokio::io::copy_buf(&mut body, &mut writer).await?; + + Ok::<(), anyhow::Error>(()) + } + }) + .collect::>(); + + while let Some(download) = downloads.next().await { + download?; + } + + Ok(temp_dir) +} + +struct BenchmarkData { + wal: Vec, + meta: BenchmarkMetadata, +} + +#[derive(Deserialize)] +struct BenchmarkMetadata { + pg_version: u32, + start_lsn: Lsn, +} + +async fn load_bench_data(path: &Utf8Path, input_size: usize) -> anyhow::Result { + eprintln!("Loading benchmark data from {:?}", path); + + let mut entries = tokio::fs::read_dir(path).await?; + let mut ordered_segment_paths = Vec::new(); + let mut metadata = None; + + while let Some(entry) = entries.next_entry().await? { + if entry.file_name() == METADATA_FILENAME { + let bytes = tokio::fs::read(entry.path()).await?; + metadata = Some( + serde_json::from_slice::(&bytes) + .context("failed to deserialize metadata.json")?, + ); + } else { + ordered_segment_paths.push(entry.path()); + } + } + + ordered_segment_paths.sort(); + + let mut buffer = Vec::new(); + for path in ordered_segment_paths { + if buffer.len() >= input_size { + break; + } + + use async_compression::tokio::bufread::ZstdDecoder; + let file = tokio::fs::File::open(path).await?; + let reader = tokio::io::BufReader::new(file); + let decoder = ZstdDecoder::new(reader); + let mut reader = tokio::io::BufReader::new(decoder); + tokio::io::copy_buf(&mut reader, &mut buffer).await?; + } + + buffer.truncate(input_size); + + Ok(BenchmarkData { + wal: buffer, + meta: metadata.unwrap(), + }) +} + +fn criterion_benchmark(c: &mut Criterion) { + const INPUT_SIZE: usize = 128 * 1024 * 1024; + + let setup_runtime = tokio::runtime::Builder::new_current_thread() + .enable_all() + .build() + .unwrap(); + + let (_temp_dir, bench_data) = setup_runtime.block_on(async move { + let cancel = CancellationToken::new(); + let client = create_s3_client().await.unwrap(); + let temp_dir = download_bench_data(client, &cancel).await.unwrap(); + let bench_data = load_bench_data(temp_dir.path(), INPUT_SIZE).await.unwrap(); + + (temp_dir, bench_data) + }); + + eprintln!( + "Benchmarking against {} MiB of WAL", + INPUT_SIZE / 1024 / 1024 + ); + + let mut group = c.benchmark_group("decode-interpret-wal"); + group.throughput(criterion::Throughput::Bytes(bench_data.wal.len() as u64)); + group.sample_size(10); + + group.bench_function("unsharded", |b| { + b.iter(|| decode_interpret_main(&bench_data, &[ShardIdentity::unsharded()])) + }); + + let eight_shards = (0..8) + .map(|i| ShardIdentity::new(ShardNumber(i), ShardCount(8), ShardStripeSize(8)).unwrap()) + .collect::>(); + + group.bench_function("8/8-shards", |b| { + b.iter(|| decode_interpret_main(&bench_data, &eight_shards)) + }); + + let four_shards = eight_shards + .into_iter() + .filter(|s| s.number.0 % 2 == 0) + .collect::>(); + group.bench_function("4/8-shards", |b| { + b.iter(|| decode_interpret_main(&bench_data, &four_shards)) + }); + + let two_shards = four_shards + .into_iter() + .filter(|s| s.number.0 % 4 == 0) + .collect::>(); + group.bench_function("2/8-shards", |b| { + b.iter(|| decode_interpret_main(&bench_data, &two_shards)) + }); +} + +fn decode_interpret_main(bench: &BenchmarkData, shards: &[ShardIdentity]) { + let r = decode_interpret(bench, shards); + if let Err(e) = r { + panic!("{e:?}"); + } +} + +fn decode_interpret(bench: &BenchmarkData, shard: &[ShardIdentity]) -> anyhow::Result<()> { + let mut decoder = WalStreamDecoder::new(bench.meta.start_lsn, bench.meta.pg_version); + let xlogoff: usize = bench.meta.start_lsn.segment_offset(WAL_SEGMENT_SIZE); + + for chunk in bench.wal[xlogoff..].chunks(MAX_SEND_SIZE) { + decoder.feed_bytes(chunk); + while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() { + assert!(lsn.is_aligned()); + let _ = InterpretedWalRecord::from_bytes_filtered( + recdata, + shard, + lsn, + bench.meta.pg_version, + ) + .unwrap(); + } + } + + Ok(()) +} +criterion_group!( + name=benches; + config=Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None))); + targets=criterion_benchmark +); +criterion_main!(benches); diff --git a/libs/wal_decoder/src/decoder.rs b/libs/wal_decoder/src/decoder.rs index aa50c62911..ebb38ceb52 100644 --- a/libs/wal_decoder/src/decoder.rs +++ b/libs/wal_decoder/src/decoder.rs @@ -1,6 +1,8 @@ //! This module contains logic for decoding and interpreting //! raw bytes which represent a raw Postgres WAL record. +use std::collections::HashMap; + use crate::models::*; use crate::serialized_batch::SerializedValueBatch; use bytes::{Buf, Bytes}; @@ -14,15 +16,15 @@ use utils::lsn::Lsn; impl InterpretedWalRecord { /// Decode and interpreted raw bytes which represent one Postgres WAL record. - /// Data blocks which do not match the provided shard identity are filtered out. + /// Data blocks which do not match any of the provided shard identities are filtered out. /// Shard 0 is a special case since it tracks all relation sizes. We only give it /// the keys that are being written as that is enough for updating relation sizes. pub fn from_bytes_filtered( buf: Bytes, - shard: &ShardIdentity, + shards: &[ShardIdentity], next_record_lsn: Lsn, pg_version: u32, - ) -> anyhow::Result { + ) -> anyhow::Result> { let mut decoded = DecodedWALRecord::default(); decode_wal_record(buf, &mut decoded, pg_version)?; let xid = decoded.xl_xid; @@ -33,43 +35,57 @@ impl InterpretedWalRecord { FlushUncommittedRecords::No }; - let metadata_record = - MetadataRecord::from_decoded_filtered(&decoded, shard, next_record_lsn, pg_version)?; - let batch = SerializedValueBatch::from_decoded_filtered( + let mut shard_records: HashMap = + HashMap::with_capacity(shards.len()); + for shard in shards { + shard_records.insert( + *shard, + InterpretedWalRecord { + metadata_record: None, + batch: SerializedValueBatch::default(), + next_record_lsn, + flush_uncommitted, + xid, + }, + ); + } + + MetadataRecord::from_decoded_filtered( + &decoded, + &mut shard_records, + next_record_lsn, + pg_version, + )?; + SerializedValueBatch::from_decoded_filtered( decoded, - shard, + &mut shard_records, next_record_lsn, pg_version, )?; - Ok(InterpretedWalRecord { - metadata_record, - batch, - next_record_lsn, - flush_uncommitted, - xid, - }) + Ok(shard_records) } } impl MetadataRecord { - /// Builds a metadata record for this WAL record, if any. + /// Populates the given `shard_records` with metadata records from this WAL record, if any, + /// discarding those belonging to other shards. /// - /// Only metadata records relevant for the given shard are emitted. Currently, most metadata + /// Only metadata records relevant for the given shards is emitted. Currently, most metadata /// records are broadcast to all shards for simplicity, but this should be improved. fn from_decoded_filtered( decoded: &DecodedWALRecord, - shard: &ShardIdentity, + shard_records: &mut HashMap, next_record_lsn: Lsn, pg_version: u32, - ) -> anyhow::Result> { + ) -> anyhow::Result<()> { // Note: this doesn't actually copy the bytes since // the [`Bytes`] type implements it via a level of indirection. let mut buf = decoded.record.clone(); buf.advance(decoded.main_data_offset); // First, generate metadata records from the decoded WAL record. - let mut metadata_record = match decoded.xl_rmid { + let metadata_record = match decoded.xl_rmid { pg_constants::RM_HEAP_ID | pg_constants::RM_HEAP2_ID => { Self::decode_heapam_record(&mut buf, decoded, pg_version)? } @@ -112,41 +128,65 @@ impl MetadataRecord { }; // Next, filter the metadata record by shard. - match metadata_record { - Some( - MetadataRecord::Heapam(HeapamRecord::ClearVmBits(ref mut clear_vm_bits)) - | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(ref mut clear_vm_bits)), - ) => { - // Route VM page updates to the shards that own them. VM pages are stored in the VM fork - // of the main relation. These are sharded and managed just like regular relation pages. - // See: https://github.com/neondatabase/neon/issues/9855 - let is_local_vm_page = |heap_blk| { - let vm_blk = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blk); - shard.is_key_local(&rel_block_to_key(clear_vm_bits.vm_rel, vm_blk)) - }; - // Send the old and new VM page updates to their respective shards. - clear_vm_bits.old_heap_blkno = clear_vm_bits - .old_heap_blkno - .filter(|&blkno| is_local_vm_page(blkno)); - clear_vm_bits.new_heap_blkno = clear_vm_bits - .new_heap_blkno - .filter(|&blkno| is_local_vm_page(blkno)); - // If neither VM page belongs to this shard, discard the record. - if clear_vm_bits.old_heap_blkno.is_none() && clear_vm_bits.new_heap_blkno.is_none() - { - metadata_record = None + for (shard, record) in shard_records.iter_mut() { + match metadata_record { + Some( + MetadataRecord::Heapam(HeapamRecord::ClearVmBits(ref clear_vm_bits)) + | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(ref clear_vm_bits)), + ) => { + // Route VM page updates to the shards that own them. VM pages are stored in the VM fork + // of the main relation. These are sharded and managed just like regular relation pages. + // See: https://github.com/neondatabase/neon/issues/9855 + let is_local_vm_page = |heap_blk| { + let vm_blk = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blk); + shard.is_key_local(&rel_block_to_key(clear_vm_bits.vm_rel, vm_blk)) + }; + // Send the old and new VM page updates to their respective shards. + let updated_old_heap_blkno = clear_vm_bits + .old_heap_blkno + .filter(|&blkno| is_local_vm_page(blkno)); + let updated_new_heap_blkno = clear_vm_bits + .new_heap_blkno + .filter(|&blkno| is_local_vm_page(blkno)); + // If neither VM page belongs to this shard, discard the record. + if updated_old_heap_blkno.is_some() || updated_new_heap_blkno.is_some() { + // Clone the record and update it for the current shard. + let mut for_shard = metadata_record.clone(); + match for_shard { + Some( + MetadataRecord::Heapam(HeapamRecord::ClearVmBits( + ref mut clear_vm_bits, + )) + | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits( + ref mut clear_vm_bits, + )), + ) => { + clear_vm_bits.old_heap_blkno = updated_old_heap_blkno; + clear_vm_bits.new_heap_blkno = updated_new_heap_blkno; + record.metadata_record = for_shard; + } + _ => { + unreachable!("for_shard is a clone of what we checked above") + } + } + } + } + Some(MetadataRecord::LogicalMessage(LogicalMessageRecord::Put(_))) => { + // Filter LogicalMessage records (AUX files) to only be stored on shard zero + if shard.is_shard_zero() { + record.metadata_record = metadata_record; + // No other shards should receive this record, so we stop traversing shards early. + break; + } + } + _ => { + // All other metadata records are sent to all shards. + record.metadata_record = metadata_record.clone(); } } - Some(MetadataRecord::LogicalMessage(LogicalMessageRecord::Put(_))) => { - // Filter LogicalMessage records (AUX files) to only be stored on shard zero - if !shard.is_shard_zero() { - metadata_record = None; - } - } - _ => {} } - Ok(metadata_record) + Ok(()) } fn decode_heapam_record( diff --git a/libs/wal_decoder/src/models.rs b/libs/wal_decoder/src/models.rs index 6576dd0eba..8bfa48faac 100644 --- a/libs/wal_decoder/src/models.rs +++ b/libs/wal_decoder/src/models.rs @@ -48,7 +48,7 @@ pub mod proto { tonic::include_proto!("interpreted_wal"); } -#[derive(Serialize, Deserialize)] +#[derive(Copy, Clone, Serialize, Deserialize)] pub enum FlushUncommittedRecords { Yes, No, @@ -107,7 +107,7 @@ impl InterpretedWalRecord { /// The interpreted part of the Postgres WAL record which requires metadata /// writes to the underlying storage engine. -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum MetadataRecord { Heapam(HeapamRecord), Neonrmgr(NeonrmgrRecord), @@ -123,12 +123,12 @@ pub enum MetadataRecord { Replorigin(ReploriginRecord), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum HeapamRecord { ClearVmBits(ClearVmBits), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct ClearVmBits { pub new_heap_blkno: Option, pub old_heap_blkno: Option, @@ -136,29 +136,29 @@ pub struct ClearVmBits { pub flags: u8, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum NeonrmgrRecord { ClearVmBits(ClearVmBits), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum SmgrRecord { Create(SmgrCreate), Truncate(XlSmgrTruncate), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct SmgrCreate { pub rel: RelTag, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum DbaseRecord { Create(DbaseCreate), Drop(DbaseDrop), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct DbaseCreate { pub db_id: Oid, pub tablespace_id: Oid, @@ -166,32 +166,32 @@ pub struct DbaseCreate { pub src_tablespace_id: Oid, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct DbaseDrop { pub db_id: Oid, pub tablespace_ids: Vec, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum ClogRecord { ZeroPage(ClogZeroPage), Truncate(ClogTruncate), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct ClogZeroPage { pub segno: u32, pub rpageno: u32, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct ClogTruncate { pub pageno: u32, pub oldest_xid: TransactionId, pub oldest_xid_db: Oid, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum XactRecord { Commit(XactCommon), Abort(XactCommon), @@ -200,7 +200,7 @@ pub enum XactRecord { Prepare(XactPrepare), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct XactCommon { pub parsed: XlXactParsedRecord, pub origin_id: u16, @@ -209,73 +209,73 @@ pub struct XactCommon { pub lsn: Lsn, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct XactPrepare { pub xl_xid: TransactionId, pub data: Bytes, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum MultiXactRecord { ZeroPage(MultiXactZeroPage), Create(XlMultiXactCreate), Truncate(XlMultiXactTruncate), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct MultiXactZeroPage { pub slru_kind: SlruKind, pub segno: u32, pub rpageno: u32, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum RelmapRecord { Update(RelmapUpdate), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct RelmapUpdate { pub update: XlRelmapUpdate, pub buf: Bytes, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum XlogRecord { Raw(RawXlogRecord), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct RawXlogRecord { pub info: u8, pub lsn: Lsn, pub buf: Bytes, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum LogicalMessageRecord { Put(PutLogicalMessage), #[cfg(feature = "testing")] Failpoint, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct PutLogicalMessage { pub path: String, pub buf: Bytes, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum StandbyRecord { RunningXacts(StandbyRunningXacts), } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub struct StandbyRunningXacts { pub oldest_running_xid: TransactionId, } -#[derive(Serialize, Deserialize)] +#[derive(Clone, Serialize, Deserialize)] pub enum ReploriginRecord { Set(XlReploriginSet), Drop(XlReploriginDrop), diff --git a/libs/wal_decoder/src/serialized_batch.rs b/libs/wal_decoder/src/serialized_batch.rs index af2b179e05..c70ff05b8e 100644 --- a/libs/wal_decoder/src/serialized_batch.rs +++ b/libs/wal_decoder/src/serialized_batch.rs @@ -5,7 +5,7 @@ //! Such batches are created from decoded PG wal records and ingested //! by the pageserver by writing directly to the ephemeral file. -use std::collections::BTreeSet; +use std::collections::{BTreeSet, HashMap}; use bytes::{Bytes, BytesMut}; use pageserver_api::key::rel_block_to_key; @@ -22,6 +22,8 @@ use utils::lsn::Lsn; use pageserver_api::key::Key; +use crate::models::InterpretedWalRecord; + static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]); /// Accompanying metadata for the batch @@ -128,7 +130,8 @@ impl Default for SerializedValueBatch { } impl SerializedValueBatch { - /// Build a batch of serialized values from a decoded PG WAL record + /// Populates the given `shard_records` with value batches from this WAL record, if any, + /// discarding those belonging to other shards. /// /// The batch will only contain values for keys targeting the specifiec /// shard. Shard 0 is a special case, where any keys that don't belong to @@ -136,21 +139,20 @@ impl SerializedValueBatch { /// but absent from the raw buffer [`SerializedValueBatch::raw`]). pub(crate) fn from_decoded_filtered( decoded: DecodedWALRecord, - shard: &ShardIdentity, + shard_records: &mut HashMap, next_record_lsn: Lsn, pg_version: u32, - ) -> anyhow::Result { - // First determine how big the buffer needs to be and allocate it up-front. + ) -> anyhow::Result<()> { + // First determine how big the buffers need to be and allocate it up-front. // This duplicates some of the work below, but it's empirically much faster. - let estimated_buffer_size = Self::estimate_buffer_size(&decoded, shard, pg_version); - let mut buf = Vec::::with_capacity(estimated_buffer_size); + for (shard, record) in shard_records.iter_mut() { + assert!(record.batch.is_empty()); + + let estimate = Self::estimate_buffer_size(&decoded, shard, pg_version); + record.batch.raw = Vec::with_capacity(estimate); + } - let mut metadata: Vec = Vec::with_capacity(decoded.blocks.len()); - let mut max_lsn: Lsn = Lsn(0); - let mut len: usize = 0; for blk in decoded.blocks.iter() { - let relative_off = buf.len() as u64; - let rel = RelTag { spcnode: blk.rnode_spcnode, dbnode: blk.rnode_dbnode, @@ -168,99 +170,98 @@ impl SerializedValueBatch { ); } - let key_is_local = shard.is_key_local(&key); + for (shard, record) in shard_records.iter_mut() { + let key_is_local = shard.is_key_local(&key); - tracing::debug!( - lsn=%next_record_lsn, - key=%key, - "ingest: shard decision {}", - if !key_is_local { "drop" } else { "keep" }, - ); + tracing::debug!( + lsn=%next_record_lsn, + key=%key, + "ingest: shard decision {}", + if !key_is_local { "drop" } else { "keep" }, + ); - if !key_is_local { - if shard.is_shard_zero() { - // Shard 0 tracks relation sizes. Although we will not store this block, we will observe - // its blkno in case it implicitly extends a relation. - metadata.push(ValueMeta::Observed(ObservedValueMeta { + if !key_is_local { + if shard.is_shard_zero() { + // Shard 0 tracks relation sizes. Although we will not store this block, we will observe + // its blkno in case it implicitly extends a relation. + record + .batch + .metadata + .push(ValueMeta::Observed(ObservedValueMeta { + key: key.to_compact(), + lsn: next_record_lsn, + })) + } + + continue; + } + + // Instead of storing full-page-image WAL record, + // it is better to store extracted image: we can skip wal-redo + // in this case. Also some FPI records may contain multiple (up to 32) pages, + // so them have to be copied multiple times. + // + let val = if Self::block_is_image(&decoded, blk, pg_version) { + // Extract page image from FPI record + let img_len = blk.bimg_len as usize; + let img_offs = blk.bimg_offset as usize; + let mut image = BytesMut::with_capacity(BLCKSZ as usize); + // TODO(vlad): skip the copy + image.extend_from_slice(&decoded.record[img_offs..img_offs + img_len]); + + if blk.hole_length != 0 { + let tail = image.split_off(blk.hole_offset as usize); + image.resize(image.len() + blk.hole_length as usize, 0u8); + image.unsplit(tail); + } + // + // Match the logic of XLogReadBufferForRedoExtended: + // The page may be uninitialized. If so, we can't set the LSN because + // that would corrupt the page. + // + if !page_is_new(&image) { + page_set_lsn(&mut image, next_record_lsn) + } + assert_eq!(image.len(), BLCKSZ as usize); + + Value::Image(image.freeze()) + } else { + Value::WalRecord(NeonWalRecord::Postgres { + will_init: blk.will_init || blk.apply_image, + rec: decoded.record.clone(), + }) + }; + + let relative_off = record.batch.raw.len() as u64; + + val.ser_into(&mut record.batch.raw) + .expect("Writing into in-memory buffer is infallible"); + + let val_ser_size = record.batch.raw.len() - relative_off as usize; + + record + .batch + .metadata + .push(ValueMeta::Serialized(SerializedValueMeta { key: key.to_compact(), lsn: next_record_lsn, - })) - } - - continue; + batch_offset: relative_off, + len: val_ser_size, + will_init: val.will_init(), + })); + record.batch.max_lsn = std::cmp::max(record.batch.max_lsn, next_record_lsn); + record.batch.len += 1; } - - // Instead of storing full-page-image WAL record, - // it is better to store extracted image: we can skip wal-redo - // in this case. Also some FPI records may contain multiple (up to 32) pages, - // so them have to be copied multiple times. - // - let val = if Self::block_is_image(&decoded, blk, pg_version) { - // Extract page image from FPI record - let img_len = blk.bimg_len as usize; - let img_offs = blk.bimg_offset as usize; - let mut image = BytesMut::with_capacity(BLCKSZ as usize); - // TODO(vlad): skip the copy - image.extend_from_slice(&decoded.record[img_offs..img_offs + img_len]); - - if blk.hole_length != 0 { - let tail = image.split_off(blk.hole_offset as usize); - image.resize(image.len() + blk.hole_length as usize, 0u8); - image.unsplit(tail); - } - // - // Match the logic of XLogReadBufferForRedoExtended: - // The page may be uninitialized. If so, we can't set the LSN because - // that would corrupt the page. - // - if !page_is_new(&image) { - page_set_lsn(&mut image, next_record_lsn) - } - assert_eq!(image.len(), BLCKSZ as usize); - - Value::Image(image.freeze()) - } else { - Value::WalRecord(NeonWalRecord::Postgres { - will_init: blk.will_init || blk.apply_image, - rec: decoded.record.clone(), - }) - }; - - val.ser_into(&mut buf) - .expect("Writing into in-memory buffer is infallible"); - - let val_ser_size = buf.len() - relative_off as usize; - - metadata.push(ValueMeta::Serialized(SerializedValueMeta { - key: key.to_compact(), - lsn: next_record_lsn, - batch_offset: relative_off, - len: val_ser_size, - will_init: val.will_init(), - })); - max_lsn = std::cmp::max(max_lsn, next_record_lsn); - len += 1; } if cfg!(any(debug_assertions, test)) { - let batch = Self { - raw: buf, - metadata, - max_lsn, - len, - }; - - batch.validate_lsn_order(); - - return Ok(batch); + // Validate that the batches are correct + for record in shard_records.values() { + record.batch.validate_lsn_order(); + } } - Ok(Self { - raw: buf, - metadata, - max_lsn, - len, - }) + Ok(()) } /// Look into the decoded PG WAL record and determine diff --git a/pageserver/src/import_datadir.rs b/pageserver/src/import_datadir.rs index c061714010..a73fa5cec8 100644 --- a/pageserver/src/import_datadir.rs +++ b/pageserver/src/import_datadir.rs @@ -278,6 +278,8 @@ async fn import_wal( let mut walingest = WalIngest::new(tline, startpoint, ctx).await?; + let shard = vec![*tline.get_shard_identity()]; + while last_lsn <= endpoint { // FIXME: assume postgresql tli 1 for now let filename = XLogFileName(1, segno, WAL_SEGMENT_SIZE); @@ -314,10 +316,12 @@ async fn import_wal( if let Some((lsn, recdata)) = waldecoder.poll_decode()? { let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - tline.get_shard_identity(), + &shard, lsn, tline.pg_version, - )?; + )? + .remove(tline.get_shard_identity()) + .unwrap(); walingest .ingest_record(interpreted, &mut modification, ctx) @@ -411,6 +415,7 @@ pub async fn import_wal_from_tar( let mut offset = start_lsn.segment_offset(WAL_SEGMENT_SIZE); let mut last_lsn = start_lsn; let mut walingest = WalIngest::new(tline, start_lsn, ctx).await?; + let shard = vec![*tline.get_shard_identity()]; // Ingest wal until end_lsn info!("importing wal until {}", end_lsn); @@ -459,10 +464,12 @@ pub async fn import_wal_from_tar( if let Some((lsn, recdata)) = waldecoder.poll_decode()? { let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - tline.get_shard_identity(), + &shard, lsn, tline.pg_version, - )?; + )? + .remove(tline.get_shard_identity()) + .unwrap(); walingest .ingest_record(interpreted, &mut modification, ctx) diff --git a/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs b/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs index 3a8796add8..129b987e57 100644 --- a/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs +++ b/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs @@ -264,6 +264,8 @@ pub(super) async fn handle_walreceiver_connection( let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx).await?; + let shard = vec![*timeline.get_shard_identity()]; + let interpreted_proto_config = match protocol { PostgresClientProtocol::Vanilla => None, PostgresClientProtocol::Interpreted { @@ -476,10 +478,12 @@ pub(super) async fn handle_walreceiver_connection( // Deserialize and interpret WAL record let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - modification.tline.get_shard_identity(), + &shard, next_record_lsn, modification.tline.pg_version, - )?; + )? + .remove(timeline.get_shard_identity()) + .unwrap(); if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes) && uncommitted_records > 0 diff --git a/pageserver/src/walingest.rs b/pageserver/src/walingest.rs index 7253af8507..ad7bcc0714 100644 --- a/pageserver/src/walingest.rs +++ b/pageserver/src/walingest.rs @@ -2163,10 +2163,12 @@ mod tests { while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() { let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - modification.tline.get_shard_identity(), + &[*modification.tline.get_shard_identity()], lsn, modification.tline.pg_version, ) + .unwrap() + .remove(modification.tline.get_shard_identity()) .unwrap(); walingest diff --git a/safekeeper/src/send_interpreted_wal.rs b/safekeeper/src/send_interpreted_wal.rs index 7d215176dd..a718c16a6a 100644 --- a/safekeeper/src/send_interpreted_wal.rs +++ b/safekeeper/src/send_interpreted_wal.rs @@ -57,6 +57,7 @@ impl InterpretedWalSender<'_, IO> { keepalive_ticker.reset(); let (tx, mut rx) = tokio::sync::mpsc::channel::(2); + let shard = vec![self.shard]; loop { tokio::select! { @@ -80,14 +81,17 @@ impl InterpretedWalSender<'_, IO> { assert!(next_record_lsn.is_aligned()); max_next_record_lsn = Some(next_record_lsn); + // Deserialize and interpret WAL record let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - &self.shard, + &shard, next_record_lsn, self.pg_version, ) - .with_context(|| "Failed to interpret WAL")?; + .with_context(|| "Failed to interpret WAL")? + .remove(&self.shard) + .unwrap(); if !interpreted.is_empty() { records.push(interpreted); From b9464865b619d35257500bfade7651e3adfc07e4 Mon Sep 17 00:00:00 2001 From: Alexander Bayandin Date: Wed, 15 Jan 2025 13:05:05 +0000 Subject: [PATCH 14/31] benchmarks: report successful runs to slack as well (#10393) ## Problem Successful `benchmarks` runs doesn't have enough visibility Ref https://neondb.slack.com/archives/C069Z2199DL/p1736868055094539 ## Summary of changes - Report both successful and failed `benchmarks` to Slack - Update `slackapi/slack-github-action` action --- .github/actionlint.yml | 1 + .github/workflows/build_and_test.yml | 25 +++++++++++-------------- 2 files changed, 12 insertions(+), 14 deletions(-) diff --git a/.github/actionlint.yml b/.github/actionlint.yml index 7a97e2ae55..aec5b4ee75 100644 --- a/.github/actionlint.yml +++ b/.github/actionlint.yml @@ -25,3 +25,4 @@ config-variables: - PGREGRESS_PG17_PROJECT_ID - SLACK_ON_CALL_QA_STAGING_STREAM - DEV_AWS_OIDC_ROLE_MANAGE_BENCHMARK_EC2_VMS_ARN + - SLACK_ON_CALL_STORAGE_STAGING_STREAM diff --git a/.github/workflows/build_and_test.yml b/.github/workflows/build_and_test.yml index cd95a5b16d..489a93f46d 100644 --- a/.github/workflows/build_and_test.yml +++ b/.github/workflows/build_and_test.yml @@ -346,25 +346,22 @@ jobs: # XXX: no coverage data handling here, since benchmarks are run on release builds, # while coverage is currently collected for the debug ones - report-benchmarks-failures: + report-benchmarks-results-to-slack: needs: [ benchmarks, create-test-report ] - if: github.ref_name == 'main' && failure() && needs.benchmarks.result == 'failure' - permissions: - id-token: write # aws-actions/configure-aws-credentials - statuses: write - contents: write - pull-requests: write + if: github.ref_name == 'main' && !cancelled() && contains(fromJSON('["success", "failure"]'), needs.benchmarks.result) runs-on: ubuntu-22.04 steps: - - uses: slackapi/slack-github-action@v1 + - uses: slackapi/slack-github-action@v2 with: - channel-id: C060CNA47S9 # on-call-staging-storage-stream - slack-message: | - Benchmarks failed on main <${{ github.event.head_commit.url }}|${{ github.sha }}> - <${{ needs.create-test-report.outputs.report-url }}|Allure report> - env: - SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }} + method: chat.postMessage + token: ${{ secrets.SLACK_BOT_TOKEN }} + payload: | + channel: "${{ vars.SLACK_ON_CALL_STORAGE_STAGING_STREAM }}" + text: | + Benchmarks on main: *${{ needs.benchmarks.result }}* + - <${{ needs.create-test-report.outputs.report-url }}|Allure report> + - <${{ github.event.head_commit.url }}|${{ github.sha }}> create-test-report: needs: [ check-permissions, build-and-test-locally, coverage-report, build-build-tools-image, benchmarks ] From 05a71c7d6a14d471dcbce9d9d27d5eed124c9947 Mon Sep 17 00:00:00 2001 From: Arseny Sher Date: Wed, 15 Jan 2025 17:16:04 +0300 Subject: [PATCH 15/31] safekeeper: add membership configuration switch endpoint (#10241) ## Problem https://github.com/neondatabase/neon/issues/9965 ## Summary of changes Add to safekeeper http endpoint to switch membership configuration. Also add it to python client for tests, and add simple test itself. --- libs/safekeeper_api/src/membership.rs | 2 + libs/safekeeper_api/src/models.rs | 15 ++++++ safekeeper/src/http/routes.rs | 28 +++++++++++ safekeeper/src/state.rs | 30 ++++++++++- safekeeper/src/timeline.rs | 20 +++++++- test_runner/fixtures/safekeeper/http.py | 43 +++++++++++++++- test_runner/regress/test_wal_acceptor.py | 64 +++++++++++++++++++++++- 7 files changed, 198 insertions(+), 4 deletions(-) diff --git a/libs/safekeeper_api/src/membership.rs b/libs/safekeeper_api/src/membership.rs index fe30204545..a39fda526f 100644 --- a/libs/safekeeper_api/src/membership.rs +++ b/libs/safekeeper_api/src/membership.rs @@ -23,6 +23,8 @@ pub const INITIAL_GENERATION: Generation = 1; pub struct SafekeeperId { pub id: NodeId, pub host: String, + /// We include here only port for computes -- that is, pg protocol tenant + /// only port, or wide pg protocol port if the former is not configured. pub pg_port: u16, } diff --git a/libs/safekeeper_api/src/models.rs b/libs/safekeeper_api/src/models.rs index ad38986357..a6f90154f4 100644 --- a/libs/safekeeper_api/src/models.rs +++ b/libs/safekeeper_api/src/models.rs @@ -175,6 +175,7 @@ pub enum WalReceiverStatus { pub struct TimelineStatus { pub tenant_id: TenantId, pub timeline_id: TimelineId, + pub mconf: Configuration, pub acceptor_state: AcceptorStateStatus, pub pg_info: ServerInfo, pub flush_lsn: Lsn, @@ -189,6 +190,20 @@ pub struct TimelineStatus { pub walreceivers: Vec, } +/// Request to switch membership configuration. +#[derive(Serialize, Deserialize)] +#[serde(transparent)] +pub struct TimelineMembershipSwitchRequest { + pub mconf: Configuration, +} + +/// In response both previous and current configuration are sent. +#[derive(Serialize, Deserialize)] +pub struct TimelineMembershipSwitchResponse { + pub previous_conf: Configuration, + pub current_conf: Configuration, +} + fn lsn_invalid() -> Lsn { Lsn::INVALID } diff --git a/safekeeper/src/http/routes.rs b/safekeeper/src/http/routes.rs index 3835d39698..5ecde4b125 100644 --- a/safekeeper/src/http/routes.rs +++ b/safekeeper/src/http/routes.rs @@ -1,4 +1,5 @@ use hyper::{Body, Request, Response, StatusCode}; +use safekeeper_api::models; use safekeeper_api::models::AcceptorStateStatus; use safekeeper_api::models::SafekeeperStatus; use safekeeper_api::models::TermSwitchApiEntry; @@ -183,6 +184,7 @@ async fn timeline_status_handler(request: Request) -> Result) -> Result, +) -> Result, ApiError> { + let ttid = TenantTimelineId::new( + parse_request_param(&request, "tenant_id")?, + parse_request_param(&request, "timeline_id")?, + ); + check_permission(&request, Some(ttid.tenant_id))?; + + let global_timelines = get_global_timelines(&request); + let tli = global_timelines.get(ttid).map_err(ApiError::from)?; + + let data: models::TimelineMembershipSwitchRequest = json_request(&mut request).await?; + let response = tli + .membership_switch(data.mconf) + .await + .map_err(ApiError::InternalServerError)?; + + json_response(StatusCode::OK, response) +} + async fn timeline_copy_handler(mut request: Request) -> Result, ApiError> { check_permission(&request, None)?; @@ -619,6 +643,10 @@ pub fn make_router( "/v1/tenant/:tenant_id/timeline/:timeline_id/snapshot/:destination_id", |r| request_span(r, timeline_snapshot_handler), ) + .post( + "/v1/tenant/:tenant_id/timeline/:timeline_id/membership", + |r| request_span(r, timeline_membership_handler), + ) .post( "/v1/tenant/:tenant_id/timeline/:source_timeline_id/copy", |r| request_span(r, timeline_copy_handler), diff --git a/safekeeper/src/state.rs b/safekeeper/src/state.rs index 1c3bb1b4dc..4d566b12a0 100644 --- a/safekeeper/src/state.rs +++ b/safekeeper/src/state.rs @@ -6,9 +6,12 @@ use std::{cmp::max, ops::Deref, time::SystemTime}; use anyhow::{bail, Result}; use postgres_ffi::WAL_SEGMENT_SIZE; use safekeeper_api::{ - membership::Configuration, models::TimelineTermBumpResponse, ServerInfo, Term, INITIAL_TERM, + membership::Configuration, + models::{TimelineMembershipSwitchResponse, TimelineTermBumpResponse}, + ServerInfo, Term, INITIAL_TERM, }; use serde::{Deserialize, Serialize}; +use tracing::info; use utils::{ id::{TenantId, TenantTimelineId, TimelineId}, lsn::Lsn, @@ -258,6 +261,31 @@ where current_term: after, }) } + + /// Switch into membership configuration `to` if it is higher than the + /// current one. + pub async fn membership_switch( + &mut self, + to: Configuration, + ) -> Result { + let before = self.mconf.clone(); + // Is switch allowed? + if to.generation <= self.mconf.generation { + info!( + "ignoring request to switch membership conf to lower {}, current conf {}", + to, self.mconf + ); + } else { + let mut state = self.start_change(); + state.mconf = to.clone(); + self.finish_change(&state).await?; + info!("switched membership conf to {} from {}", to, before); + } + Ok(TimelineMembershipSwitchResponse { + previous_conf: before, + current_conf: self.mconf.clone(), + }) + } } impl Deref for TimelineState diff --git a/safekeeper/src/timeline.rs b/safekeeper/src/timeline.rs index 36860a0da2..2882391074 100644 --- a/safekeeper/src/timeline.rs +++ b/safekeeper/src/timeline.rs @@ -4,7 +4,10 @@ use anyhow::{anyhow, bail, Result}; use camino::{Utf8Path, Utf8PathBuf}; use remote_storage::RemotePath; -use safekeeper_api::models::{PeerInfo, TimelineTermBumpResponse}; +use safekeeper_api::membership::Configuration; +use safekeeper_api::models::{ + PeerInfo, TimelineMembershipSwitchResponse, TimelineTermBumpResponse, +}; use safekeeper_api::Term; use tokio::fs::{self}; use tokio_util::sync::CancellationToken; @@ -188,6 +191,13 @@ impl StateSK { self.state_mut().term_bump(to).await } + pub async fn membership_switch( + &mut self, + to: Configuration, + ) -> Result { + self.state_mut().membership_switch(to).await + } + /// Close open WAL files to release FDs. fn close_wal_store(&mut self) { if let StateSK::Loaded(sk) = self { @@ -768,6 +778,14 @@ impl Timeline { state.sk.term_bump(to).await } + pub async fn membership_switch( + self: &Arc, + to: Configuration, + ) -> Result { + let mut state = self.write_shared_state().await; + state.sk.membership_switch(to).await + } + /// Guts of [`Self::wal_residence_guard`] and [`Self::try_wal_residence_guard`] async fn do_wal_residence_guard( self: &Arc, diff --git a/test_runner/fixtures/safekeeper/http.py b/test_runner/fixtures/safekeeper/http.py index 4826cae3ee..493ce7334e 100644 --- a/test_runner/fixtures/safekeeper/http.py +++ b/test_runner/fixtures/safekeeper/http.py @@ -25,6 +25,7 @@ class Walreceiver: @dataclass class SafekeeperTimelineStatus: + mconf: Configuration | None term: int last_log_term: int pg_version: int # Not exactly a PgVersion, safekeeper returns version as int, for example 150002 for 15.2 @@ -73,7 +74,7 @@ class TermBumpResponse: class SafekeeperId: id: int host: str - pg_port: str + pg_port: int @dataclass @@ -82,6 +83,16 @@ class Configuration: members: list[SafekeeperId] new_members: list[SafekeeperId] | None + @classmethod + def from_json(cls, d: dict[str, Any]) -> Configuration: + generation = d["generation"] + members = d["members"] + new_members = d.get("new_members") + return Configuration(generation, members, new_members) + + def to_json(self) -> str: + return json.dumps(self, cls=EnhancedJSONEncoder) + @dataclass class TimelineCreateRequest: @@ -97,6 +108,18 @@ class TimelineCreateRequest: return json.dumps(self, cls=EnhancedJSONEncoder) +@dataclass +class TimelineMembershipSwitchResponse: + previous_conf: Configuration + current_conf: Configuration + + @classmethod + def from_json(cls, d: dict[str, Any]) -> TimelineMembershipSwitchResponse: + previous_conf = Configuration.from_json(d["previous_conf"]) + current_conf = Configuration.from_json(d["current_conf"]) + return TimelineMembershipSwitchResponse(previous_conf, current_conf) + + class SafekeeperHttpClient(requests.Session, MetricsGetter): HTTPError = requests.HTTPError @@ -170,7 +193,10 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter): res.raise_for_status() resj = res.json() walreceivers = [Walreceiver(wr["conn_id"], wr["status"]) for wr in resj["walreceivers"]] + # It is always normally not None, it is allowed only to make forward compat tests happy. + mconf = Configuration.from_json(resj["mconf"]) if "mconf" in resj else None return SafekeeperTimelineStatus( + mconf=mconf, term=resj["acceptor_state"]["term"], last_log_term=resj["acceptor_state"]["epoch"], pg_version=resj["pg_info"]["pg_version"], @@ -196,6 +222,11 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter): def get_commit_lsn(self, tenant_id: TenantId, timeline_id: TimelineId) -> Lsn: return self.timeline_status(tenant_id, timeline_id).commit_lsn + # Get timeline membership configuration. + def get_membership(self, tenant_id: TenantId, timeline_id: TimelineId) -> Configuration: + # make mypy happy + return self.timeline_status(tenant_id, timeline_id).mconf # type: ignore + # only_local doesn't remove segments in the remote storage. def timeline_delete( self, tenant_id: TenantId, timeline_id: TimelineId, only_local: bool = False @@ -242,6 +273,16 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter): assert isinstance(res_json, dict) return res_json + def membership_switch( + self, tenant_id: TenantId, timeline_id: TimelineId, to: Configuration + ) -> TimelineMembershipSwitchResponse: + res = self.post( + f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/membership", + data=to.to_json(), + ) + res.raise_for_status() + return TimelineMembershipSwitchResponse.from_json(res.json()) + def copy_timeline(self, tenant_id: TenantId, timeline_id: TimelineId, body: dict[str, Any]): res = self.post( f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/copy", diff --git a/test_runner/regress/test_wal_acceptor.py b/test_runner/regress/test_wal_acceptor.py index d39c6a6b5b..2b6a267bdf 100644 --- a/test_runner/regress/test_wal_acceptor.py +++ b/test_runner/regress/test_wal_acceptor.py @@ -48,7 +48,12 @@ from fixtures.remote_storage import ( default_remote_storage, s3_storage, ) -from fixtures.safekeeper.http import Configuration, SafekeeperHttpClient, TimelineCreateRequest +from fixtures.safekeeper.http import ( + Configuration, + SafekeeperHttpClient, + SafekeeperId, + TimelineCreateRequest, +) from fixtures.safekeeper.utils import wait_walreceivers_absent from fixtures.utils import ( PropagatingThread, @@ -2243,6 +2248,63 @@ def test_pull_timeline_while_evicted(neon_env_builder: NeonEnvBuilder): wait_until(unevicted_on_dest, interval=0.1, timeout=1.0) +# Basic test for http API membership related calls: create timeline and switch +# configuration. Normally these are called by storage controller, but this +# allows to test them separately. +@run_only_on_default_postgres("tests only safekeeper API") +def test_membership_api(neon_env_builder: NeonEnvBuilder): + neon_env_builder.num_safekeepers = 1 + env = neon_env_builder.init_start() + + tenant_id = env.initial_tenant + timeline_id = env.initial_timeline + + sk = env.safekeepers[0] + http_cli = sk.http_client() + + sk_id_1 = SafekeeperId(env.safekeepers[0].id, "localhost", sk.port.pg_tenant_only) + sk_id_2 = SafekeeperId(11, "localhost", 5434) # just a mock + + # Request to switch before timeline creation should fail. + init_conf = Configuration(generation=1, members=[sk_id_1], new_members=None) + with pytest.raises(requests.exceptions.HTTPError): + http_cli.membership_switch(tenant_id, timeline_id, init_conf) + + # Create timeline. + create_r = TimelineCreateRequest( + tenant_id, timeline_id, init_conf, 150002, Lsn("0/1000000"), commit_lsn=None + ) + log.info(f"sending {create_r.to_json()}") + http_cli.timeline_create(create_r) + + # Switch into some conf. + joint_conf = Configuration(generation=4, members=[sk_id_1], new_members=[sk_id_2]) + resp = http_cli.membership_switch(tenant_id, timeline_id, joint_conf) + log.info(f"joint switch resp: {resp}") + assert resp.previous_conf.generation == 1 + assert resp.current_conf.generation == 4 + + # Restart sk, conf should be preserved. + sk.stop().start() + after_restart = http_cli.get_membership(tenant_id, timeline_id) + log.info(f"conf after restart: {after_restart}") + assert after_restart.generation == 4 + + # Switch into disjoint conf. + non_joint = Configuration(generation=5, members=[sk_id_2], new_members=None) + resp = http_cli.membership_switch(tenant_id, timeline_id, non_joint) + log.info(f"non joint switch resp: {resp}") + assert resp.previous_conf.generation == 4 + assert resp.current_conf.generation == 5 + + # Switch request to lower conf should be ignored. + lower_conf = Configuration(generation=3, members=[], new_members=None) + resp = http_cli.membership_switch(tenant_id, timeline_id, lower_conf) + log.info(f"lower switch resp: {resp}") + assert resp.previous_conf.generation == 5 + assert resp.current_conf.generation == 5 + + # In this test we check for excessive START_REPLICATION and START_WAL_PUSH queries # when compute is active, but there are no writes to the timeline. In that case # pageserver should maintain a single connection to safekeeper and don't attempt From 3e529f124f45f20f0e10948b4b4a8ba881b59b06 Mon Sep 17 00:00:00 2001 From: Tristan Partin Date: Wed, 15 Jan 2025 09:29:52 -0600 Subject: [PATCH 16/31] Remove leading slashes when downloading remote files (#10396) Signed-off-by: Tristan Partin --- vendor/postgres-v14 | 2 +- vendor/postgres-v15 | 2 +- vendor/postgres-v16 | 2 +- vendor/postgres-v17 | 2 +- vendor/revisions.json | 8 ++++---- 5 files changed, 8 insertions(+), 8 deletions(-) diff --git a/vendor/postgres-v14 b/vendor/postgres-v14 index 210a0ba3af..46082f2088 160000 --- a/vendor/postgres-v14 +++ b/vendor/postgres-v14 @@ -1 +1 @@ -Subproject commit 210a0ba3afd8134ea910b203f274b165bd4f05d7 +Subproject commit 46082f20884f087a2d974b33ac65d63af26142bd diff --git a/vendor/postgres-v15 b/vendor/postgres-v15 index d3141e17a7..dd0b28d6fb 160000 --- a/vendor/postgres-v15 +++ b/vendor/postgres-v15 @@ -1 +1 @@ -Subproject commit d3141e17a7155e3d07c8deba4a10c748a29ba1e6 +Subproject commit dd0b28d6fbad39e227f3b77296fcca879af8b3a9 diff --git a/vendor/postgres-v16 b/vendor/postgres-v16 index f63b141cfb..d674efd776 160000 --- a/vendor/postgres-v16 +++ b/vendor/postgres-v16 @@ -1 +1 @@ -Subproject commit f63b141cfb0c813725a6b2574049565bff643018 +Subproject commit d674efd776f59d78e8fa1535bd2f95c3e6984fca diff --git a/vendor/postgres-v17 b/vendor/postgres-v17 index 0f8da73ed0..a8dd6e779d 160000 --- a/vendor/postgres-v17 +++ b/vendor/postgres-v17 @@ -1 +1 @@ -Subproject commit 0f8da73ed08d4fc4ee58cccea008c75bfb20baa8 +Subproject commit a8dd6e779dde907778006adb436b557ad652fb97 diff --git a/vendor/revisions.json b/vendor/revisions.json index b4d57ab709..c899dbaa5a 100644 --- a/vendor/revisions.json +++ b/vendor/revisions.json @@ -1,18 +1,18 @@ { "v17": [ "17.2", - "0f8da73ed08d4fc4ee58cccea008c75bfb20baa8" + "a8dd6e779dde907778006adb436b557ad652fb97" ], "v16": [ "16.6", - "f63b141cfb0c813725a6b2574049565bff643018" + "d674efd776f59d78e8fa1535bd2f95c3e6984fca" ], "v15": [ "15.10", - "d3141e17a7155e3d07c8deba4a10c748a29ba1e6" + "dd0b28d6fbad39e227f3b77296fcca879af8b3a9" ], "v14": [ "14.15", - "210a0ba3afd8134ea910b203f274b165bd4f05d7" + "46082f20884f087a2d974b33ac65d63af26142bd" ] } From dbebede7bf5ff0cefd303f266781457b7472d070 Mon Sep 17 00:00:00 2001 From: Vlad Lazar Date: Wed, 15 Jan 2025 15:33:54 +0000 Subject: [PATCH 17/31] safekeeper: fan out from single wal reader to multiple shards (#10190) ## Problem Safekeepers currently decode and interpret WAL for each shard separately. This is wasteful in terms of CPU memory usage - we've seen this in profiles. ## Summary of changes Fan-out interpreted WAL to multiple shards. The basic is that wal decoding and interpretation happens in a separate tokio task and senders attach to it. Senders only receive batches concerning their shard and only past the Lsn they've last seen. Fan-out is gated behind the `wal_reader_fanout` safekeeper flag (disabled by default for now). When fan-out is enabled, it might be desirable to control the absolute delta between the current position and a new shard's desired position (i.e. how far behind or ahead a shard may be). `max_delta_for_fanout` is a new optional safekeeper flag which dictates whether to create a new WAL reader or attach to the existing one. By default, this behaviour is disabled. Let's consider enabling it if we spot the need for it in the field. ## Testing Tests passed [here](https://github.com/neondatabase/neon/pull/10301) with wal reader fanout enabled as of https://github.com/neondatabase/neon/pull/10190/commits/34f6a717182c431847bbd5b7828fd0f89027b2be. Related: https://github.com/neondatabase/neon/issues/9337 Epic: https://github.com/neondatabase/neon/issues/9329 --- Cargo.lock | 3 + libs/safekeeper_api/Cargo.toml | 1 + libs/safekeeper_api/src/models.rs | 20 +- libs/wal_decoder/src/models.rs | 2 +- libs/wal_decoder/src/serialized_batch.rs | 8 +- safekeeper/Cargo.toml | 3 + safekeeper/src/bin/safekeeper.rs | 9 + safekeeper/src/http/routes.rs | 2 +- safekeeper/src/lib.rs | 4 + safekeeper/src/metrics.rs | 32 +- safekeeper/src/send_interpreted_wal.rs | 765 ++++++++++++++++-- safekeeper/src/send_wal.rs | 354 ++++++-- safekeeper/src/test_utils.rs | 65 +- safekeeper/src/timeline.rs | 14 +- safekeeper/src/wal_reader_stream.rs | 396 ++++++--- .../tests/walproposer_sim/safekeeper.rs | 2 + 16 files changed, 1410 insertions(+), 270 deletions(-) diff --git a/Cargo.lock b/Cargo.lock index 0669899617..afe16ff848 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -5655,6 +5655,7 @@ dependencies = [ "crc32c", "criterion", "desim", + "env_logger 0.10.2", "fail", "futures", "hex", @@ -5683,6 +5684,7 @@ dependencies = [ "serde", "serde_json", "sha2", + "smallvec", "storage_broker", "strum", "strum_macros", @@ -5709,6 +5711,7 @@ version = "0.1.0" dependencies = [ "anyhow", "const_format", + "pageserver_api", "postgres_ffi", "pq_proto", "serde", diff --git a/libs/safekeeper_api/Cargo.toml b/libs/safekeeper_api/Cargo.toml index 7652c3d413..6b72ace019 100644 --- a/libs/safekeeper_api/Cargo.toml +++ b/libs/safekeeper_api/Cargo.toml @@ -13,3 +13,4 @@ postgres_ffi.workspace = true pq_proto.workspace = true tokio.workspace = true utils.workspace = true +pageserver_api.workspace = true diff --git a/libs/safekeeper_api/src/models.rs b/libs/safekeeper_api/src/models.rs index a6f90154f4..b5fa903820 100644 --- a/libs/safekeeper_api/src/models.rs +++ b/libs/safekeeper_api/src/models.rs @@ -1,5 +1,6 @@ //! Types used in safekeeper http API. Many of them are also reused internally. +use pageserver_api::shard::ShardIdentity; use postgres_ffi::TimestampTz; use serde::{Deserialize, Serialize}; use std::net::SocketAddr; @@ -146,7 +147,13 @@ pub type ConnectionId = u32; /// Serialize is used only for json'ing in API response. Also used internally. #[derive(Debug, Clone, Serialize, Deserialize)] -pub struct WalSenderState { +pub enum WalSenderState { + Vanilla(VanillaWalSenderState), + Interpreted(InterpretedWalSenderState), +} + +#[derive(Debug, Clone, Serialize, Deserialize)] +pub struct VanillaWalSenderState { pub ttid: TenantTimelineId, pub addr: SocketAddr, pub conn_id: ConnectionId, @@ -155,6 +162,17 @@ pub struct WalSenderState { pub feedback: ReplicationFeedback, } +#[derive(Debug, Clone, Serialize, Deserialize)] +pub struct InterpretedWalSenderState { + pub ttid: TenantTimelineId, + pub shard: ShardIdentity, + pub addr: SocketAddr, + pub conn_id: ConnectionId, + // postgres application_name + pub appname: Option, + pub feedback: ReplicationFeedback, +} + #[derive(Debug, Clone, Serialize, Deserialize)] pub struct WalReceiverState { /// None means it is recovery initiated by us (this safekeeper). diff --git a/libs/wal_decoder/src/models.rs b/libs/wal_decoder/src/models.rs index 8bfa48faac..c2f9125b21 100644 --- a/libs/wal_decoder/src/models.rs +++ b/libs/wal_decoder/src/models.rs @@ -64,7 +64,7 @@ pub struct InterpretedWalRecords { } /// An interpreted Postgres WAL record, ready to be handled by the pageserver -#[derive(Serialize, Deserialize)] +#[derive(Serialize, Deserialize, Clone)] pub struct InterpretedWalRecord { /// Optional metadata record - may cause writes to metadata keys /// in the storage engine diff --git a/libs/wal_decoder/src/serialized_batch.rs b/libs/wal_decoder/src/serialized_batch.rs index c70ff05b8e..d76f75f51f 100644 --- a/libs/wal_decoder/src/serialized_batch.rs +++ b/libs/wal_decoder/src/serialized_batch.rs @@ -32,7 +32,7 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]); /// relation sizes. In the case of "observed" values, we only need to know /// the key and LSN, so two types of metadata are supported to save on network /// bandwidth. -#[derive(Serialize, Deserialize)] +#[derive(Serialize, Deserialize, Clone)] pub enum ValueMeta { Serialized(SerializedValueMeta), Observed(ObservedValueMeta), @@ -79,7 +79,7 @@ impl PartialEq for OrderedValueMeta { impl Eq for OrderedValueMeta {} /// Metadata for a [`Value`] serialized into the batch. -#[derive(Serialize, Deserialize)] +#[derive(Serialize, Deserialize, Clone)] pub struct SerializedValueMeta { pub key: CompactKey, pub lsn: Lsn, @@ -91,14 +91,14 @@ pub struct SerializedValueMeta { } /// Metadata for a [`Value`] observed by the batch -#[derive(Serialize, Deserialize)] +#[derive(Serialize, Deserialize, Clone)] pub struct ObservedValueMeta { pub key: CompactKey, pub lsn: Lsn, } /// Batch of serialized [`Value`]s. -#[derive(Serialize, Deserialize)] +#[derive(Serialize, Deserialize, Clone)] pub struct SerializedValueBatch { /// [`Value`]s serialized in EphemeralFile's native format, /// ready for disk write by the pageserver diff --git a/safekeeper/Cargo.toml b/safekeeper/Cargo.toml index 3ebb7097f2..0eb511f1cc 100644 --- a/safekeeper/Cargo.toml +++ b/safekeeper/Cargo.toml @@ -26,6 +26,7 @@ hex.workspace = true humantime.workspace = true http.workspace = true hyper0.workspace = true +itertools.workspace = true futures.workspace = true once_cell.workspace = true parking_lot.workspace = true @@ -39,6 +40,7 @@ scopeguard.workspace = true reqwest = { workspace = true, features = ["json"] } serde.workspace = true serde_json.workspace = true +smallvec.workspace = true strum.workspace = true strum_macros.workspace = true thiserror.workspace = true @@ -63,6 +65,7 @@ storage_broker.workspace = true tokio-stream.workspace = true utils.workspace = true wal_decoder.workspace = true +env_logger.workspace = true workspace_hack.workspace = true diff --git a/safekeeper/src/bin/safekeeper.rs b/safekeeper/src/bin/safekeeper.rs index bc7af02185..6cc53e0d23 100644 --- a/safekeeper/src/bin/safekeeper.rs +++ b/safekeeper/src/bin/safekeeper.rs @@ -207,6 +207,13 @@ struct Args { /// Also defines interval for eviction retries. #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_EVICTION_MIN_RESIDENT)] eviction_min_resident: Duration, + /// Enable fanning out WAL to different shards from the same reader + #[arg(long)] + wal_reader_fanout: bool, + /// Only fan out the WAL reader if the absoulte delta between the new requested position + /// and the current position of the reader is smaller than this value. + #[arg(long)] + max_delta_for_fanout: Option, } // Like PathBufValueParser, but allows empty string. @@ -370,6 +377,8 @@ async fn main() -> anyhow::Result<()> { control_file_save_interval: args.control_file_save_interval, partial_backup_concurrency: args.partial_backup_concurrency, eviction_min_resident: args.eviction_min_resident, + wal_reader_fanout: args.wal_reader_fanout, + max_delta_for_fanout: args.max_delta_for_fanout, }); // initialize sentry if SENTRY_DSN is provided diff --git a/safekeeper/src/http/routes.rs b/safekeeper/src/http/routes.rs index 5ecde4b125..4b9fb9eb67 100644 --- a/safekeeper/src/http/routes.rs +++ b/safekeeper/src/http/routes.rs @@ -195,7 +195,7 @@ async fn timeline_status_handler(request: Request) -> Result, } impl SafeKeeperConf { @@ -150,6 +152,8 @@ impl SafeKeeperConf { control_file_save_interval: Duration::from_secs(1), partial_backup_concurrency: 1, eviction_min_resident: Duration::ZERO, + wal_reader_fanout: false, + max_delta_for_fanout: None, } } } diff --git a/safekeeper/src/metrics.rs b/safekeeper/src/metrics.rs index 5883f402c7..3ea9e3d674 100644 --- a/safekeeper/src/metrics.rs +++ b/safekeeper/src/metrics.rs @@ -12,9 +12,9 @@ use metrics::{ pow2_buckets, proto::MetricFamily, register_histogram, register_histogram_vec, register_int_counter, register_int_counter_pair, - register_int_counter_pair_vec, register_int_counter_vec, register_int_gauge, Gauge, GaugeVec, - Histogram, HistogramVec, IntCounter, IntCounterPair, IntCounterPairVec, IntCounterVec, - IntGauge, IntGaugeVec, DISK_FSYNC_SECONDS_BUCKETS, + register_int_counter_pair_vec, register_int_counter_vec, register_int_gauge, + register_int_gauge_vec, Gauge, GaugeVec, Histogram, HistogramVec, IntCounter, IntCounterPair, + IntCounterPairVec, IntCounterVec, IntGauge, IntGaugeVec, DISK_FSYNC_SECONDS_BUCKETS, }; use once_cell::sync::Lazy; use postgres_ffi::XLogSegNo; @@ -211,6 +211,14 @@ pub static WAL_RECEIVERS: Lazy = Lazy::new(|| { ) .expect("Failed to register safekeeper_wal_receivers") }); +pub static WAL_READERS: Lazy = Lazy::new(|| { + register_int_gauge_vec!( + "safekeeper_wal_readers", + "Number of active WAL readers (may serve pageservers or other safekeepers)", + &["kind", "target"] + ) + .expect("Failed to register safekeeper_wal_receivers") +}); pub static WAL_RECEIVER_QUEUE_DEPTH: Lazy = Lazy::new(|| { // Use powers of two buckets, but add a bucket at 0 and the max queue size to track empty and // full queues respectively. @@ -443,6 +451,7 @@ pub struct FullTimelineInfo { pub timeline_is_active: bool, pub num_computes: u32, pub last_removed_segno: XLogSegNo, + pub interpreted_wal_reader_tasks: usize, pub epoch_start_lsn: Lsn, pub mem_state: TimelineMemState, @@ -472,6 +481,7 @@ pub struct TimelineCollector { disk_usage: GenericGaugeVec, acceptor_term: GenericGaugeVec, written_wal_bytes: GenericGaugeVec, + interpreted_wal_reader_tasks: GenericGaugeVec, written_wal_seconds: GaugeVec, flushed_wal_seconds: GaugeVec, collect_timeline_metrics: Gauge, @@ -670,6 +680,16 @@ impl TimelineCollector { .unwrap(); descs.extend(active_timelines_count.desc().into_iter().cloned()); + let interpreted_wal_reader_tasks = GenericGaugeVec::new( + Opts::new( + "safekeeper_interpreted_wal_reader_tasks", + "Number of active interpreted wal reader tasks, grouped by timeline", + ), + &["tenant_id", "timeline_id"], + ) + .unwrap(); + descs.extend(interpreted_wal_reader_tasks.desc().into_iter().cloned()); + TimelineCollector { global_timelines, descs, @@ -693,6 +713,7 @@ impl TimelineCollector { collect_timeline_metrics, timelines_count, active_timelines_count, + interpreted_wal_reader_tasks, } } } @@ -721,6 +742,7 @@ impl Collector for TimelineCollector { self.disk_usage.reset(); self.acceptor_term.reset(); self.written_wal_bytes.reset(); + self.interpreted_wal_reader_tasks.reset(); self.written_wal_seconds.reset(); self.flushed_wal_seconds.reset(); @@ -782,6 +804,9 @@ impl Collector for TimelineCollector { self.written_wal_bytes .with_label_values(labels) .set(tli.wal_storage.write_wal_bytes); + self.interpreted_wal_reader_tasks + .with_label_values(labels) + .set(tli.interpreted_wal_reader_tasks as u64); self.written_wal_seconds .with_label_values(labels) .set(tli.wal_storage.write_wal_seconds); @@ -834,6 +859,7 @@ impl Collector for TimelineCollector { mfs.extend(self.disk_usage.collect()); mfs.extend(self.acceptor_term.collect()); mfs.extend(self.written_wal_bytes.collect()); + mfs.extend(self.interpreted_wal_reader_tasks.collect()); mfs.extend(self.written_wal_seconds.collect()); mfs.extend(self.flushed_wal_seconds.collect()); diff --git a/safekeeper/src/send_interpreted_wal.rs b/safekeeper/src/send_interpreted_wal.rs index a718c16a6a..ea09ce364d 100644 --- a/safekeeper/src/send_interpreted_wal.rs +++ b/safekeeper/src/send_interpreted_wal.rs @@ -1,100 +1,330 @@ +use std::collections::HashMap; +use std::fmt::Display; +use std::sync::Arc; use std::time::Duration; -use anyhow::Context; +use anyhow::{anyhow, Context}; +use futures::future::Either; use futures::StreamExt; use pageserver_api::shard::ShardIdentity; use postgres_backend::{CopyStreamHandlerEnd, PostgresBackend}; -use postgres_ffi::MAX_SEND_SIZE; +use postgres_ffi::waldecoder::WalDecodeError; use postgres_ffi::{get_current_timestamp, waldecoder::WalStreamDecoder}; use pq_proto::{BeMessage, InterpretedWalRecordsBody, WalSndKeepAlive}; use tokio::io::{AsyncRead, AsyncWrite}; +use tokio::sync::mpsc::error::SendError; +use tokio::task::JoinHandle; use tokio::time::MissedTickBehavior; +use tracing::{info_span, Instrument}; use utils::lsn::Lsn; use utils::postgres_client::Compression; use utils::postgres_client::InterpretedFormat; use wal_decoder::models::{InterpretedWalRecord, InterpretedWalRecords}; use wal_decoder::wire_format::ToWireFormat; -use crate::send_wal::EndWatchView; -use crate::wal_reader_stream::{WalBytes, WalReaderStreamBuilder}; +use crate::metrics::WAL_READERS; +use crate::send_wal::{EndWatchView, WalSenderGuard}; +use crate::timeline::WalResidentTimeline; +use crate::wal_reader_stream::{StreamingWalReader, WalBytes}; -/// Shard-aware interpreted record sender. -/// This is used for sending WAL to the pageserver. Said WAL -/// is pre-interpreted and filtered for the shard. -pub(crate) struct InterpretedWalSender<'a, IO> { - pub(crate) format: InterpretedFormat, - pub(crate) compression: Option, - pub(crate) pgb: &'a mut PostgresBackend, - pub(crate) wal_stream_builder: WalReaderStreamBuilder, - pub(crate) end_watch_view: EndWatchView, - pub(crate) shard: ShardIdentity, - pub(crate) pg_version: u32, - pub(crate) appname: Option, +/// Identifier used to differentiate between senders of the same +/// shard. +/// +/// In the steady state there's only one, but two pageservers may +/// temporarily have the same shard attached and attempt to ingest +/// WAL for it. See also [`ShardSenderId`]. +#[derive(Hash, Eq, PartialEq, Copy, Clone)] +struct SenderId(u8); + +impl SenderId { + fn first() -> Self { + SenderId(0) + } + + fn next(&self) -> Self { + SenderId(self.0.checked_add(1).expect("few senders")) + } } -struct Batch { +#[derive(Hash, Eq, PartialEq)] +struct ShardSenderId { + shard: ShardIdentity, + sender_id: SenderId, +} + +impl Display for ShardSenderId { + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + write!(f, "{}{}", self.sender_id.0, self.shard.shard_slug()) + } +} + +impl ShardSenderId { + fn new(shard: ShardIdentity, sender_id: SenderId) -> Self { + ShardSenderId { shard, sender_id } + } + + fn shard(&self) -> ShardIdentity { + self.shard + } +} + +/// Shard-aware fan-out interpreted record reader. +/// Reads WAL from disk, decodes it, intepretets it, and sends +/// it to any [`InterpretedWalSender`] connected to it. +/// Each [`InterpretedWalSender`] corresponds to one shard +/// and gets interpreted records concerning that shard only. +pub(crate) struct InterpretedWalReader { + wal_stream: StreamingWalReader, + shard_senders: HashMap>, + shard_notification_rx: Option>, + state: Arc>, + pg_version: u32, +} + +/// A handle for [`InterpretedWalReader`] which allows for interacting with it +/// when it runs as a separate tokio task. +#[derive(Debug)] +pub(crate) struct InterpretedWalReaderHandle { + join_handle: JoinHandle>, + state: Arc>, + shard_notification_tx: tokio::sync::mpsc::UnboundedSender, +} + +struct ShardSenderState { + sender_id: SenderId, + tx: tokio::sync::mpsc::Sender, + next_record_lsn: Lsn, +} + +/// State of [`InterpretedWalReader`] visible outside of the task running it. +#[derive(Debug)] +pub(crate) enum InterpretedWalReaderState { + Running { current_position: Lsn }, + Done, +} + +pub(crate) struct Batch { wal_end_lsn: Lsn, available_wal_end_lsn: Lsn, records: InterpretedWalRecords, } -impl InterpretedWalSender<'_, IO> { - /// Send interpreted WAL to a receiver. - /// Stops when an error occurs or the receiver is caught up and there's no active compute. - /// - /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ? - /// convenience. - pub(crate) async fn run(self) -> Result<(), CopyStreamHandlerEnd> { - let mut wal_position = self.wal_stream_builder.start_pos(); - let mut wal_decoder = - WalStreamDecoder::new(self.wal_stream_builder.start_pos(), self.pg_version); +#[derive(thiserror::Error, Debug)] +pub enum InterpretedWalReaderError { + /// Handler initiates the end of streaming. + #[error("decode error: {0}")] + Decode(#[from] WalDecodeError), + #[error("read or interpret error: {0}")] + ReadOrInterpret(#[from] anyhow::Error), + #[error("wal stream closed")] + WalStreamClosed, +} - let stream = self.wal_stream_builder.build(MAX_SEND_SIZE).await?; - let mut stream = std::pin::pin!(stream); +impl InterpretedWalReaderState { + fn current_position(&self) -> Option { + match self { + InterpretedWalReaderState::Running { + current_position, .. + } => Some(*current_position), + InterpretedWalReaderState::Done => None, + } + } +} - let mut keepalive_ticker = tokio::time::interval(Duration::from_secs(1)); - keepalive_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip); - keepalive_ticker.reset(); +pub(crate) struct AttachShardNotification { + shard_id: ShardIdentity, + sender: tokio::sync::mpsc::Sender, + start_pos: Lsn, +} - let (tx, mut rx) = tokio::sync::mpsc::channel::(2); - let shard = vec![self.shard]; +impl InterpretedWalReader { + /// Spawn the reader in a separate tokio task and return a handle + pub(crate) fn spawn( + wal_stream: StreamingWalReader, + start_pos: Lsn, + tx: tokio::sync::mpsc::Sender, + shard: ShardIdentity, + pg_version: u32, + appname: &Option, + ) -> InterpretedWalReaderHandle { + let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running { + current_position: start_pos, + })); + + let (shard_notification_tx, shard_notification_rx) = tokio::sync::mpsc::unbounded_channel(); + + let reader = InterpretedWalReader { + wal_stream, + shard_senders: HashMap::from([( + shard, + smallvec::smallvec![ShardSenderState { + sender_id: SenderId::first(), + tx, + next_record_lsn: start_pos, + }], + )]), + shard_notification_rx: Some(shard_notification_rx), + state: state.clone(), + pg_version, + }; + + let metric = WAL_READERS + .get_metric_with_label_values(&["task", appname.as_deref().unwrap_or("safekeeper")]) + .unwrap(); + + let join_handle = tokio::task::spawn( + async move { + metric.inc(); + scopeguard::defer! { + metric.dec(); + } + + let res = reader.run_impl(start_pos).await; + if let Err(ref err) = res { + tracing::error!("Task finished with error: {err}"); + } + res + } + .instrument(info_span!("interpreted wal reader")), + ); + + InterpretedWalReaderHandle { + join_handle, + state, + shard_notification_tx, + } + } + + /// Construct the reader without spawning anything + /// Callers should drive the future returned by [`Self::run`]. + pub(crate) fn new( + wal_stream: StreamingWalReader, + start_pos: Lsn, + tx: tokio::sync::mpsc::Sender, + shard: ShardIdentity, + pg_version: u32, + ) -> InterpretedWalReader { + let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running { + current_position: start_pos, + })); + + InterpretedWalReader { + wal_stream, + shard_senders: HashMap::from([( + shard, + smallvec::smallvec![ShardSenderState { + sender_id: SenderId::first(), + tx, + next_record_lsn: start_pos, + }], + )]), + shard_notification_rx: None, + state: state.clone(), + pg_version, + } + } + + /// Entry point for future (polling) based wal reader. + pub(crate) async fn run( + self, + start_pos: Lsn, + appname: &Option, + ) -> Result<(), CopyStreamHandlerEnd> { + let metric = WAL_READERS + .get_metric_with_label_values(&["future", appname.as_deref().unwrap_or("safekeeper")]) + .unwrap(); + + metric.inc(); + scopeguard::defer! { + metric.dec(); + } + + let res = self.run_impl(start_pos).await; + if let Err(err) = res { + tracing::error!("Interpreted wal reader encountered error: {err}"); + } else { + tracing::info!("Interpreted wal reader exiting"); + } + + Err(CopyStreamHandlerEnd::Other(anyhow!( + "interpreted wal reader finished" + ))) + } + + /// Send interpreted WAL to one or more [`InterpretedWalSender`]s + /// Stops when an error is encountered or when the [`InterpretedWalReaderHandle`] + /// goes out of scope. + async fn run_impl(mut self, start_pos: Lsn) -> Result<(), InterpretedWalReaderError> { + let defer_state = self.state.clone(); + scopeguard::defer! { + *defer_state.write().unwrap() = InterpretedWalReaderState::Done; + } + + let mut wal_decoder = WalStreamDecoder::new(start_pos, self.pg_version); loop { tokio::select! { - // Get some WAL from the stream and then: decode, interpret and push it down the - // pipeline. - wal = stream.next(), if tx.capacity() > 0 => { - let WalBytes { wal, wal_start_lsn: _, wal_end_lsn, available_wal_end_lsn } = match wal { - Some(some) => some?, - None => { break; } + // Main branch for reading WAL and forwarding it + wal_or_reset = self.wal_stream.next() => { + let wal = wal_or_reset.map(|wor| wor.get_wal().expect("reset handled in select branch below")); + let WalBytes { + wal, + wal_start_lsn: _, + wal_end_lsn, + available_wal_end_lsn, + } = match wal { + Some(some) => some.map_err(InterpretedWalReaderError::ReadOrInterpret)?, + None => { + // [`StreamingWalReader::next`] is an endless stream of WAL. + // It shouldn't ever finish unless it panicked or became internally + // inconsistent. + return Result::Err(InterpretedWalReaderError::WalStreamClosed); + } }; - wal_position = wal_end_lsn; wal_decoder.feed_bytes(&wal); - let mut records = Vec::new(); + // Deserialize and interpret WAL records from this batch of WAL. + // Interpreted records for each shard are collected separately. + let shard_ids = self.shard_senders.keys().copied().collect::>(); + let mut records_by_sender: HashMap> = HashMap::new(); let mut max_next_record_lsn = None; - while let Some((next_record_lsn, recdata)) = wal_decoder - .poll_decode() - .with_context(|| "Failed to decode WAL")? + while let Some((next_record_lsn, recdata)) = wal_decoder.poll_decode()? { assert!(next_record_lsn.is_aligned()); max_next_record_lsn = Some(next_record_lsn); - - // Deserialize and interpret WAL record let interpreted = InterpretedWalRecord::from_bytes_filtered( recdata, - &shard, + &shard_ids, next_record_lsn, self.pg_version, ) - .with_context(|| "Failed to interpret WAL")? - .remove(&self.shard) - .unwrap(); + .with_context(|| "Failed to interpret WAL")?; - if !interpreted.is_empty() { - records.push(interpreted); + for (shard, record) in interpreted { + if record.is_empty() { + continue; + } + + let mut states_iter = self.shard_senders + .get(&shard) + .expect("keys collected above") + .iter() + .filter(|state| record.next_record_lsn > state.next_record_lsn) + .peekable(); + while let Some(state) = states_iter.next() { + let shard_sender_id = ShardSenderId::new(shard, state.sender_id); + + // The most commont case is one sender per shard. Peek and break to avoid the + // clone in that situation. + if states_iter.peek().is_none() { + records_by_sender.entry(shard_sender_id).or_default().push(record); + break; + } else { + records_by_sender.entry(shard_sender_id).or_default().push(record.clone()); + } + } } } @@ -103,20 +333,170 @@ impl InterpretedWalSender<'_, IO> { None => { continue; } }; - let batch = InterpretedWalRecords { - records, - next_record_lsn: Some(max_next_record_lsn), - }; + // Update the current position such that new receivers can decide + // whether to attach to us or spawn a new WAL reader. + match &mut *self.state.write().unwrap() { + InterpretedWalReaderState::Running { current_position, .. } => { + *current_position = max_next_record_lsn; + }, + InterpretedWalReaderState::Done => { + unreachable!() + } + } - tx.send(Batch {wal_end_lsn, available_wal_end_lsn, records: batch}).await.unwrap(); + // Send interpreted records downstream. Anything that has already been seen + // by a shard is filtered out. + let mut shard_senders_to_remove = Vec::new(); + for (shard, states) in &mut self.shard_senders { + for state in states { + if max_next_record_lsn <= state.next_record_lsn { + continue; + } + + let shard_sender_id = ShardSenderId::new(*shard, state.sender_id); + let records = records_by_sender.remove(&shard_sender_id).unwrap_or_default(); + + let batch = InterpretedWalRecords { + records, + next_record_lsn: Some(max_next_record_lsn), + }; + + let res = state.tx.send(Batch { + wal_end_lsn, + available_wal_end_lsn, + records: batch, + }).await; + + if res.is_err() { + shard_senders_to_remove.push(shard_sender_id); + } else { + state.next_record_lsn = max_next_record_lsn; + } + } + } + + // Clean up any shard senders that have dropped out. + // This is inefficient, but such events are rare (connection to PS termination) + // and the number of subscriptions on the same shards very small (only one + // for the steady state). + for to_remove in shard_senders_to_remove { + let shard_senders = self.shard_senders.get_mut(&to_remove.shard()).expect("saw it above"); + if let Some(idx) = shard_senders.iter().position(|s| s.sender_id == to_remove.sender_id) { + shard_senders.remove(idx); + tracing::info!("Removed shard sender {}", to_remove); + } + + if shard_senders.is_empty() { + self.shard_senders.remove(&to_remove.shard()); + } + } }, - // For a previously interpreted batch, serialize it and push it down the wire. - batch = rx.recv() => { + // Listen for new shards that want to attach to this reader. + // If the reader is not running as a task, then this is not supported + // (see the pending branch below). + notification = match self.shard_notification_rx.as_mut() { + Some(rx) => Either::Left(rx.recv()), + None => Either::Right(std::future::pending()) + } => { + if let Some(n) = notification { + let AttachShardNotification { shard_id, sender, start_pos } = n; + + // Update internal and external state, then reset the WAL stream + // if required. + let senders = self.shard_senders.entry(shard_id).or_default(); + let new_sender_id = match senders.last() { + Some(sender) => sender.sender_id.next(), + None => SenderId::first() + }; + + senders.push(ShardSenderState { sender_id: new_sender_id, tx: sender, next_record_lsn: start_pos}); + let current_pos = self.state.read().unwrap().current_position().unwrap(); + if start_pos < current_pos { + self.wal_stream.reset(start_pos).await; + wal_decoder = WalStreamDecoder::new(start_pos, self.pg_version); + } + + tracing::info!( + "Added shard sender {} with start_pos={} current_pos={}", + ShardSenderId::new(shard_id, new_sender_id), start_pos, current_pos + ); + } + } + } + } + } +} + +impl InterpretedWalReaderHandle { + /// Fan-out the reader by attaching a new shard to it + pub(crate) fn fanout( + &self, + shard_id: ShardIdentity, + sender: tokio::sync::mpsc::Sender, + start_pos: Lsn, + ) -> Result<(), SendError> { + self.shard_notification_tx.send(AttachShardNotification { + shard_id, + sender, + start_pos, + }) + } + + /// Get the current WAL position of the reader + pub(crate) fn current_position(&self) -> Option { + self.state.read().unwrap().current_position() + } + + pub(crate) fn abort(&self) { + self.join_handle.abort() + } +} + +impl Drop for InterpretedWalReaderHandle { + fn drop(&mut self) { + tracing::info!("Aborting interpreted wal reader"); + self.abort() + } +} + +pub(crate) struct InterpretedWalSender<'a, IO> { + pub(crate) format: InterpretedFormat, + pub(crate) compression: Option, + pub(crate) appname: Option, + + pub(crate) tli: WalResidentTimeline, + pub(crate) start_lsn: Lsn, + + pub(crate) pgb: &'a mut PostgresBackend, + pub(crate) end_watch_view: EndWatchView, + pub(crate) wal_sender_guard: Arc, + pub(crate) rx: tokio::sync::mpsc::Receiver, +} + +impl InterpretedWalSender<'_, IO> { + /// Send interpreted WAL records over the network. + /// Also manages keep-alives if nothing was sent for a while. + pub(crate) async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> { + let mut keepalive_ticker = tokio::time::interval(Duration::from_secs(1)); + keepalive_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip); + keepalive_ticker.reset(); + + let mut wal_position = self.start_lsn; + + loop { + tokio::select! { + batch = self.rx.recv() => { let batch = match batch { Some(b) => b, - None => { break; } + None => { + return Result::Err( + CopyStreamHandlerEnd::Other(anyhow!("Interpreted WAL reader exited early")) + ); + } }; + wal_position = batch.wal_end_lsn; + let buf = batch .records .to_wire(self.format, self.compression) @@ -136,7 +516,21 @@ impl InterpretedWalSender<'_, IO> { })).await?; } // Send a periodic keep alive when the connection has been idle for a while. + // Since we've been idle, also check if we can stop streaming. _ = keepalive_ticker.tick() => { + if let Some(remote_consistent_lsn) = self.wal_sender_guard + .walsenders() + .get_ws_remote_consistent_lsn(self.wal_sender_guard.id()) + { + if self.tli.should_walsender_stop(remote_consistent_lsn).await { + // Stop streaming if the receivers are caught up and + // there's no active compute. This causes the loop in + // [`crate::send_interpreted_wal::InterpretedWalSender::run`] + // to exit and terminate the WAL stream. + break; + } + } + self.pgb .write_message(&BeMessage::KeepAlive(WalSndKeepAlive { wal_end: self.end_watch_view.get().0, @@ -144,14 +538,259 @@ impl InterpretedWalSender<'_, IO> { request_reply: true, })) .await?; - } + }, } } - // The loop above ends when the receiver is caught up and there's no more WAL to send. Err(CopyStreamHandlerEnd::ServerInitiated(format!( "ending streaming to {:?} at {}, receiver is caughtup and there is no computes", self.appname, wal_position, ))) } } +#[cfg(test)] +mod tests { + use std::{collections::HashMap, str::FromStr, time::Duration}; + + use pageserver_api::shard::{ShardIdentity, ShardStripeSize}; + use postgres_ffi::MAX_SEND_SIZE; + use tokio::sync::mpsc::error::TryRecvError; + use utils::{ + id::{NodeId, TenantTimelineId}, + lsn::Lsn, + shard::{ShardCount, ShardNumber}, + }; + + use crate::{ + send_interpreted_wal::{Batch, InterpretedWalReader}, + test_utils::Env, + wal_reader_stream::StreamingWalReader, + }; + + #[tokio::test] + async fn test_interpreted_wal_reader_fanout() { + let _ = env_logger::builder().is_test(true).try_init(); + + const SIZE: usize = 8 * 1024; + const MSG_COUNT: usize = 200; + const PG_VERSION: u32 = 17; + const SHARD_COUNT: u8 = 2; + + let start_lsn = Lsn::from_str("0/149FD18").unwrap(); + let env = Env::new(true).unwrap(); + let tli = env + .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn) + .await + .unwrap(); + + let resident_tli = tli.wal_residence_guard().await.unwrap(); + let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT) + .await + .unwrap(); + let end_pos = end_watch.get(); + + tracing::info!("Doing first round of reads ..."); + + let streaming_wal_reader = StreamingWalReader::new( + resident_tli, + None, + start_lsn, + end_pos, + end_watch, + MAX_SEND_SIZE, + ); + + let shard_0 = ShardIdentity::new( + ShardNumber(0), + ShardCount(SHARD_COUNT), + ShardStripeSize::default(), + ) + .unwrap(); + + let shard_1 = ShardIdentity::new( + ShardNumber(1), + ShardCount(SHARD_COUNT), + ShardStripeSize::default(), + ) + .unwrap(); + + let mut shards = HashMap::new(); + + for shard_number in 0..SHARD_COUNT { + let shard_id = ShardIdentity::new( + ShardNumber(shard_number), + ShardCount(SHARD_COUNT), + ShardStripeSize::default(), + ) + .unwrap(); + let (tx, rx) = tokio::sync::mpsc::channel::(MSG_COUNT * 2); + shards.insert(shard_id, (Some(tx), Some(rx))); + } + + let shard_0_tx = shards.get_mut(&shard_0).unwrap().0.take().unwrap(); + let mut shard_0_rx = shards.get_mut(&shard_0).unwrap().1.take().unwrap(); + + let handle = InterpretedWalReader::spawn( + streaming_wal_reader, + start_lsn, + shard_0_tx, + shard_0, + PG_VERSION, + &Some("pageserver".to_string()), + ); + + tracing::info!("Reading all WAL with only shard 0 attached ..."); + + let mut shard_0_interpreted_records = Vec::new(); + while let Some(batch) = shard_0_rx.recv().await { + shard_0_interpreted_records.push(batch.records); + if batch.wal_end_lsn == batch.available_wal_end_lsn { + break; + } + } + + let shard_1_tx = shards.get_mut(&shard_1).unwrap().0.take().unwrap(); + let mut shard_1_rx = shards.get_mut(&shard_1).unwrap().1.take().unwrap(); + + tracing::info!("Attaching shard 1 to the reader at start of WAL"); + handle.fanout(shard_1, shard_1_tx, start_lsn).unwrap(); + + tracing::info!("Reading all WAL with shard 0 and shard 1 attached ..."); + + let mut shard_1_interpreted_records = Vec::new(); + while let Some(batch) = shard_1_rx.recv().await { + shard_1_interpreted_records.push(batch.records); + if batch.wal_end_lsn == batch.available_wal_end_lsn { + break; + } + } + + // This test uses logical messages. Those only go to shard 0. Check that the + // filtering worked and shard 1 did not get any. + assert!(shard_1_interpreted_records + .iter() + .all(|recs| recs.records.is_empty())); + + // Shard 0 should not receive anything more since the reader is + // going through wal that it has already processed. + let res = shard_0_rx.try_recv(); + if let Ok(ref ok) = res { + tracing::error!( + "Shard 0 received batch: wal_end_lsn={} available_wal_end_lsn={}", + ok.wal_end_lsn, + ok.available_wal_end_lsn + ); + } + assert!(matches!(res, Err(TryRecvError::Empty))); + + // Check that the next records lsns received by the two shards match up. + let shard_0_next_lsns = shard_0_interpreted_records + .iter() + .map(|recs| recs.next_record_lsn) + .collect::>(); + let shard_1_next_lsns = shard_1_interpreted_records + .iter() + .map(|recs| recs.next_record_lsn) + .collect::>(); + assert_eq!(shard_0_next_lsns, shard_1_next_lsns); + + handle.abort(); + let mut done = false; + for _ in 0..5 { + if handle.current_position().is_none() { + done = true; + break; + } + tokio::time::sleep(Duration::from_millis(1)).await; + } + + assert!(done); + } + + #[tokio::test] + async fn test_interpreted_wal_reader_same_shard_fanout() { + let _ = env_logger::builder().is_test(true).try_init(); + + const SIZE: usize = 8 * 1024; + const MSG_COUNT: usize = 200; + const PG_VERSION: u32 = 17; + const SHARD_COUNT: u8 = 2; + const ATTACHED_SHARDS: u8 = 4; + + let start_lsn = Lsn::from_str("0/149FD18").unwrap(); + let env = Env::new(true).unwrap(); + let tli = env + .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn) + .await + .unwrap(); + + let resident_tli = tli.wal_residence_guard().await.unwrap(); + let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT) + .await + .unwrap(); + let end_pos = end_watch.get(); + + let streaming_wal_reader = StreamingWalReader::new( + resident_tli, + None, + start_lsn, + end_pos, + end_watch, + MAX_SEND_SIZE, + ); + + let shard_0 = ShardIdentity::new( + ShardNumber(0), + ShardCount(SHARD_COUNT), + ShardStripeSize::default(), + ) + .unwrap(); + + let (tx, rx) = tokio::sync::mpsc::channel::(MSG_COUNT * 2); + let mut batch_receivers = vec![rx]; + + let handle = InterpretedWalReader::spawn( + streaming_wal_reader, + start_lsn, + tx, + shard_0, + PG_VERSION, + &Some("pageserver".to_string()), + ); + + for _ in 0..(ATTACHED_SHARDS - 1) { + let (tx, rx) = tokio::sync::mpsc::channel::(MSG_COUNT * 2); + handle.fanout(shard_0, tx, start_lsn).unwrap(); + batch_receivers.push(rx); + } + + loop { + let batch = batch_receivers.first_mut().unwrap().recv().await.unwrap(); + for rx in batch_receivers.iter_mut().skip(1) { + let other_batch = rx.recv().await.unwrap(); + + assert_eq!(batch.wal_end_lsn, other_batch.wal_end_lsn); + assert_eq!( + batch.available_wal_end_lsn, + other_batch.available_wal_end_lsn + ); + } + + if batch.wal_end_lsn == batch.available_wal_end_lsn { + break; + } + } + + handle.abort(); + let mut done = false; + for _ in 0..5 { + if handle.current_position().is_none() { + done = true; + break; + } + tokio::time::sleep(Duration::from_millis(1)).await; + } + + assert!(done); + } +} diff --git a/safekeeper/src/send_wal.rs b/safekeeper/src/send_wal.rs index 8463221998..4a4a74a0fd 100644 --- a/safekeeper/src/send_wal.rs +++ b/safekeeper/src/send_wal.rs @@ -2,16 +2,18 @@ //! with the "START_REPLICATION" message, and registry of walsenders. use crate::handler::SafekeeperPostgresHandler; -use crate::metrics::RECEIVED_PS_FEEDBACKS; +use crate::metrics::{RECEIVED_PS_FEEDBACKS, WAL_READERS}; use crate::receive_wal::WalReceivers; use crate::safekeeper::TermLsn; -use crate::send_interpreted_wal::InterpretedWalSender; +use crate::send_interpreted_wal::{ + Batch, InterpretedWalReader, InterpretedWalReaderHandle, InterpretedWalSender, +}; use crate::timeline::WalResidentTimeline; -use crate::wal_reader_stream::WalReaderStreamBuilder; +use crate::wal_reader_stream::StreamingWalReader; use crate::wal_storage::WalReader; use anyhow::{bail, Context as AnyhowContext}; use bytes::Bytes; -use futures::future::Either; +use futures::FutureExt; use parking_lot::Mutex; use postgres_backend::PostgresBackend; use postgres_backend::{CopyStreamHandlerEnd, PostgresBackendReader, QueryError}; @@ -19,16 +21,16 @@ use postgres_ffi::get_current_timestamp; use postgres_ffi::{TimestampTz, MAX_SEND_SIZE}; use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody}; use safekeeper_api::models::{ - ConnectionId, HotStandbyFeedback, ReplicationFeedback, StandbyFeedback, StandbyReply, - WalSenderState, INVALID_FULL_TRANSACTION_ID, + HotStandbyFeedback, ReplicationFeedback, StandbyFeedback, StandbyReply, + INVALID_FULL_TRANSACTION_ID, }; use safekeeper_api::Term; use tokio::io::{AsyncRead, AsyncWrite}; use utils::failpoint_support; -use utils::id::TenantTimelineId; use utils::pageserver_feedback::PageserverFeedback; use utils::postgres_client::PostgresClientProtocol; +use itertools::Itertools; use std::cmp::{max, min}; use std::net::SocketAddr; use std::sync::Arc; @@ -50,6 +52,12 @@ pub struct WalSenders { walreceivers: Arc, } +pub struct WalSendersTimelineMetricValues { + pub ps_feedback_counter: u64, + pub last_ps_feedback: PageserverFeedback, + pub interpreted_wal_reader_tasks: usize, +} + impl WalSenders { pub fn new(walreceivers: Arc) -> Arc { Arc::new(WalSenders { @@ -60,21 +68,8 @@ impl WalSenders { /// Register new walsender. Returned guard provides access to the slot and /// automatically deregisters in Drop. - fn register( - self: &Arc, - ttid: TenantTimelineId, - addr: SocketAddr, - conn_id: ConnectionId, - appname: Option, - ) -> WalSenderGuard { + fn register(self: &Arc, walsender_state: WalSenderState) -> WalSenderGuard { let slots = &mut self.mutex.lock().slots; - let walsender_state = WalSenderState { - ttid, - addr, - conn_id, - appname, - feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()), - }; // find empty slot or create new one let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) { slots[pos] = Some(walsender_state); @@ -90,9 +85,79 @@ impl WalSenders { } } + fn create_or_update_interpreted_reader< + FUp: FnOnce(&Arc) -> anyhow::Result<()>, + FNew: FnOnce() -> InterpretedWalReaderHandle, + >( + self: &Arc, + id: WalSenderId, + start_pos: Lsn, + max_delta_for_fanout: Option, + update: FUp, + create: FNew, + ) -> anyhow::Result<()> { + let state = &mut self.mutex.lock(); + + let mut selected_interpreted_reader = None; + for slot in state.slots.iter().flatten() { + if let WalSenderState::Interpreted(slot_state) = slot { + if let Some(ref interpreted_reader) = slot_state.interpreted_wal_reader { + let select = match (interpreted_reader.current_position(), max_delta_for_fanout) + { + (Some(pos), Some(max_delta)) => { + let delta = pos.0.abs_diff(start_pos.0); + delta <= max_delta + } + // Reader is not active + (None, _) => false, + // Gating fanout by max delta is disabled. + // Attach to any active reader. + (_, None) => true, + }; + + if select { + selected_interpreted_reader = Some(interpreted_reader.clone()); + break; + } + } + } + } + + let slot = state.get_slot_mut(id); + let slot_state = match slot { + WalSenderState::Interpreted(s) => s, + WalSenderState::Vanilla(_) => unreachable!(), + }; + + let selected_or_new = match selected_interpreted_reader { + Some(selected) => { + update(&selected)?; + selected + } + None => Arc::new(create()), + }; + + slot_state.interpreted_wal_reader = Some(selected_or_new); + + Ok(()) + } + /// Get state of all walsenders. - pub fn get_all(self: &Arc) -> Vec { - self.mutex.lock().slots.iter().flatten().cloned().collect() + pub fn get_all_public(self: &Arc) -> Vec { + self.mutex + .lock() + .slots + .iter() + .flatten() + .map(|state| match state { + WalSenderState::Vanilla(s) => { + safekeeper_api::models::WalSenderState::Vanilla(s.clone()) + } + WalSenderState::Interpreted(s) => { + safekeeper_api::models::WalSenderState::Interpreted(s.public_state.clone()) + } + }) + .collect() } /// Get LSN of the most lagging pageserver receiver. Return None if there are no @@ -103,7 +168,7 @@ impl WalSenders { .slots .iter() .flatten() - .filter_map(|s| match s.feedback { + .filter_map(|s| match s.get_feedback() { ReplicationFeedback::Pageserver(feedback) => Some(feedback.last_received_lsn), ReplicationFeedback::Standby(_) => None, }) @@ -111,9 +176,25 @@ impl WalSenders { } /// Returns total counter of pageserver feedbacks received and last feedback. - pub fn get_ps_feedback_stats(self: &Arc) -> (u64, PageserverFeedback) { + pub fn info_for_metrics(self: &Arc) -> WalSendersTimelineMetricValues { let shared = self.mutex.lock(); - (shared.ps_feedback_counter, shared.last_ps_feedback) + + let interpreted_wal_reader_tasks = shared + .slots + .iter() + .filter_map(|ss| match ss { + Some(WalSenderState::Interpreted(int)) => int.interpreted_wal_reader.as_ref(), + Some(WalSenderState::Vanilla(_)) => None, + None => None, + }) + .unique_by(|reader| Arc::as_ptr(reader)) + .count(); + + WalSendersTimelineMetricValues { + ps_feedback_counter: shared.ps_feedback_counter, + last_ps_feedback: shared.last_ps_feedback, + interpreted_wal_reader_tasks, + } } /// Get aggregated hot standby feedback (we send it to compute). @@ -124,7 +205,7 @@ impl WalSenders { /// Record new pageserver feedback, update aggregated values. fn record_ps_feedback(self: &Arc, id: WalSenderId, feedback: &PageserverFeedback) { let mut shared = self.mutex.lock(); - shared.get_slot_mut(id).feedback = ReplicationFeedback::Pageserver(*feedback); + *shared.get_slot_mut(id).get_mut_feedback() = ReplicationFeedback::Pageserver(*feedback); shared.last_ps_feedback = *feedback; shared.ps_feedback_counter += 1; drop(shared); @@ -143,10 +224,10 @@ impl WalSenders { "Record standby reply: ts={} apply_lsn={}", reply.reply_ts, reply.apply_lsn ); - match &mut slot.feedback { + match &mut slot.get_mut_feedback() { ReplicationFeedback::Standby(sf) => sf.reply = *reply, ReplicationFeedback::Pageserver(_) => { - slot.feedback = ReplicationFeedback::Standby(StandbyFeedback { + *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback { reply: *reply, hs_feedback: HotStandbyFeedback::empty(), }) @@ -158,10 +239,10 @@ impl WalSenders { fn record_hs_feedback(self: &Arc, id: WalSenderId, feedback: &HotStandbyFeedback) { let mut shared = self.mutex.lock(); let slot = shared.get_slot_mut(id); - match &mut slot.feedback { + match &mut slot.get_mut_feedback() { ReplicationFeedback::Standby(sf) => sf.hs_feedback = *feedback, ReplicationFeedback::Pageserver(_) => { - slot.feedback = ReplicationFeedback::Standby(StandbyFeedback { + *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback { reply: StandbyReply::empty(), hs_feedback: *feedback, }) @@ -175,7 +256,7 @@ impl WalSenders { pub fn get_ws_remote_consistent_lsn(self: &Arc, id: WalSenderId) -> Option { let shared = self.mutex.lock(); let slot = shared.get_slot(id); - match slot.feedback { + match slot.get_feedback() { ReplicationFeedback::Pageserver(feedback) => Some(feedback.remote_consistent_lsn), _ => None, } @@ -199,6 +280,47 @@ struct WalSendersShared { slots: Vec>, } +/// Safekeeper internal definitions of wal sender state +/// +/// As opposed to [`safekeeper_api::models::WalSenderState`] these struct may +/// include state that we don not wish to expose to the public api. +#[derive(Debug, Clone)] +pub(crate) enum WalSenderState { + Vanilla(VanillaWalSenderInternalState), + Interpreted(InterpretedWalSenderInternalState), +} + +type VanillaWalSenderInternalState = safekeeper_api::models::VanillaWalSenderState; + +#[derive(Debug, Clone)] +pub(crate) struct InterpretedWalSenderInternalState { + public_state: safekeeper_api::models::InterpretedWalSenderState, + interpreted_wal_reader: Option>, +} + +impl WalSenderState { + fn get_addr(&self) -> &SocketAddr { + match self { + WalSenderState::Vanilla(state) => &state.addr, + WalSenderState::Interpreted(state) => &state.public_state.addr, + } + } + + fn get_feedback(&self) -> &ReplicationFeedback { + match self { + WalSenderState::Vanilla(state) => &state.feedback, + WalSenderState::Interpreted(state) => &state.public_state.feedback, + } + } + + fn get_mut_feedback(&mut self) -> &mut ReplicationFeedback { + match self { + WalSenderState::Vanilla(state) => &mut state.feedback, + WalSenderState::Interpreted(state) => &mut state.public_state.feedback, + } + } +} + impl WalSendersShared { fn new() -> Self { WalSendersShared { @@ -225,7 +347,7 @@ impl WalSendersShared { let mut agg = HotStandbyFeedback::empty(); let mut reply_agg = StandbyReply::empty(); for ws_state in self.slots.iter().flatten() { - if let ReplicationFeedback::Standby(standby_feedback) = ws_state.feedback { + if let ReplicationFeedback::Standby(standby_feedback) = ws_state.get_feedback() { let hs_feedback = standby_feedback.hs_feedback; // doing Option math like op1.iter().chain(op2.iter()).min() // would be nicer, but we serialize/deserialize this struct @@ -317,7 +439,7 @@ impl SafekeeperPostgresHandler { /// Wrapper around handle_start_replication_guts handling result. Error is /// handled here while we're still in walsender ttid span; with API /// extension, this can probably be moved into postgres_backend. - pub async fn handle_start_replication( + pub async fn handle_start_replication( &mut self, pgb: &mut PostgresBackend, start_pos: Lsn, @@ -342,7 +464,7 @@ impl SafekeeperPostgresHandler { Ok(()) } - pub async fn handle_start_replication_guts( + pub async fn handle_start_replication_guts( &mut self, pgb: &mut PostgresBackend, start_pos: Lsn, @@ -352,12 +474,30 @@ impl SafekeeperPostgresHandler { let appname = self.appname.clone(); // Use a guard object to remove our entry from the timeline when we are done. - let ws_guard = Arc::new(tli.get_walsenders().register( - self.ttid, - *pgb.get_peer_addr(), - self.conn_id, - self.appname.clone(), - )); + let ws_guard = match self.protocol() { + PostgresClientProtocol::Vanilla => Arc::new(tli.get_walsenders().register( + WalSenderState::Vanilla(VanillaWalSenderInternalState { + ttid: self.ttid, + addr: *pgb.get_peer_addr(), + conn_id: self.conn_id, + appname: self.appname.clone(), + feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()), + }), + )), + PostgresClientProtocol::Interpreted { .. } => Arc::new(tli.get_walsenders().register( + WalSenderState::Interpreted(InterpretedWalSenderInternalState { + public_state: safekeeper_api::models::InterpretedWalSenderState { + ttid: self.ttid, + shard: self.shard.unwrap(), + addr: *pgb.get_peer_addr(), + conn_id: self.conn_id, + appname: self.appname.clone(), + feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()), + }, + interpreted_wal_reader: None, + }), + )), + }; // Walsender can operate in one of two modes which we select by // application_name: give only committed WAL (used by pageserver) or all @@ -403,7 +543,7 @@ impl SafekeeperPostgresHandler { pgb, // should succeed since we're already holding another guard tli: tli.wal_residence_guard().await?, - appname, + appname: appname.clone(), start_pos, end_pos, term, @@ -413,7 +553,7 @@ impl SafekeeperPostgresHandler { send_buf: vec![0u8; MAX_SEND_SIZE], }; - Either::Left(sender.run()) + FutureExt::boxed(sender.run()) } PostgresClientProtocol::Interpreted { format, @@ -421,27 +561,96 @@ impl SafekeeperPostgresHandler { } => { let pg_version = tli.tli.get_state().await.1.server.pg_version / 10000; let end_watch_view = end_watch.view(); - let wal_stream_builder = WalReaderStreamBuilder { - tli: tli.wal_residence_guard().await?, - start_pos, - end_pos, - term, - end_watch, - wal_sender_guard: ws_guard.clone(), - }; + let wal_residence_guard = tli.wal_residence_guard().await?; + let (tx, rx) = tokio::sync::mpsc::channel::(2); + let shard = self.shard.unwrap(); - let sender = InterpretedWalSender { - format, - compression, - pgb, - wal_stream_builder, - end_watch_view, - shard: self.shard.unwrap(), - pg_version, - appname, - }; + if self.conf.wal_reader_fanout && !shard.is_unsharded() { + let ws_id = ws_guard.id(); + ws_guard.walsenders().create_or_update_interpreted_reader( + ws_id, + start_pos, + self.conf.max_delta_for_fanout, + { + let tx = tx.clone(); + |reader| { + tracing::info!( + "Fanning out interpreted wal reader at {}", + start_pos + ); + reader + .fanout(shard, tx, start_pos) + .with_context(|| "Failed to fan out reader") + } + }, + || { + tracing::info!("Spawning interpreted wal reader at {}", start_pos); - Either::Right(sender.run()) + let wal_stream = StreamingWalReader::new( + wal_residence_guard, + term, + start_pos, + end_pos, + end_watch, + MAX_SEND_SIZE, + ); + + InterpretedWalReader::spawn( + wal_stream, start_pos, tx, shard, pg_version, &appname, + ) + }, + )?; + + let sender = InterpretedWalSender { + format, + compression, + appname, + tli: tli.wal_residence_guard().await?, + start_lsn: start_pos, + pgb, + end_watch_view, + wal_sender_guard: ws_guard.clone(), + rx, + }; + + FutureExt::boxed(sender.run()) + } else { + let wal_reader = StreamingWalReader::new( + wal_residence_guard, + term, + start_pos, + end_pos, + end_watch, + MAX_SEND_SIZE, + ); + + let reader = + InterpretedWalReader::new(wal_reader, start_pos, tx, shard, pg_version); + + let sender = InterpretedWalSender { + format, + compression, + appname: appname.clone(), + tli: tli.wal_residence_guard().await?, + start_lsn: start_pos, + pgb, + end_watch_view, + wal_sender_guard: ws_guard.clone(), + rx, + }; + + FutureExt::boxed(async move { + // Sender returns an Err on all code paths. + // If the sender finishes first, we will drop the reader future. + // If the reader finishes first, the sender will finish too since + // the wal sender has dropped. + let res = tokio::try_join!(sender.run(), reader.run(start_pos, &appname)); + match res.map(|_| ()) { + Ok(_) => unreachable!("sender finishes with Err by convention"), + err_res => err_res, + } + }) + } } }; @@ -470,7 +679,8 @@ impl SafekeeperPostgresHandler { .clone(); info!( "finished streaming to {}, feedback={:?}", - ws_state.addr, ws_state.feedback, + ws_state.get_addr(), + ws_state.get_feedback(), ); // Join pg backend back. @@ -578,6 +788,18 @@ impl WalSender<'_, IO> { /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ? /// convenience. async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> { + let metric = WAL_READERS + .get_metric_with_label_values(&[ + "future", + self.appname.as_deref().unwrap_or("safekeeper"), + ]) + .unwrap(); + + metric.inc(); + scopeguard::defer! { + metric.dec(); + } + loop { // Wait for the next portion if it is not there yet, or just // update our end of WAL available for sending value, we @@ -813,7 +1035,7 @@ impl ReplyReader { #[cfg(test)] mod tests { use safekeeper_api::models::FullTransactionId; - use utils::id::{TenantId, TimelineId}; + use utils::id::{TenantId, TenantTimelineId, TimelineId}; use super::*; @@ -830,13 +1052,13 @@ mod tests { // add to wss specified feedback setting other fields to dummy values fn push_feedback(wss: &mut WalSendersShared, feedback: ReplicationFeedback) { - let walsender_state = WalSenderState { + let walsender_state = WalSenderState::Vanilla(VanillaWalSenderInternalState { ttid: mock_ttid(), addr: mock_addr(), conn_id: 1, appname: None, feedback, - }; + }); wss.slots.push(Some(walsender_state)) } diff --git a/safekeeper/src/test_utils.rs b/safekeeper/src/test_utils.rs index c40a8bae5a..4e851c5b3d 100644 --- a/safekeeper/src/test_utils.rs +++ b/safekeeper/src/test_utils.rs @@ -1,13 +1,19 @@ use std::sync::Arc; use crate::rate_limit::RateLimiter; -use crate::safekeeper::{ProposerAcceptorMessage, ProposerElected, SafeKeeper, TermHistory}; +use crate::receive_wal::WalAcceptor; +use crate::safekeeper::{ + AcceptorProposerMessage, AppendRequest, AppendRequestHeader, ProposerAcceptorMessage, + ProposerElected, SafeKeeper, TermHistory, +}; +use crate::send_wal::EndWatch; use crate::state::{TimelinePersistentState, TimelineState}; use crate::timeline::{get_timeline_dir, SharedState, StateSK, Timeline}; use crate::timelines_set::TimelinesSet; use crate::wal_backup::remote_timeline_path; -use crate::{control_file, wal_storage, SafeKeeperConf}; +use crate::{control_file, receive_wal, wal_storage, SafeKeeperConf}; use camino_tempfile::Utf8TempDir; +use postgres_ffi::v17::wal_generator::{LogicalMessageGenerator, WalGenerator}; use tokio::fs::create_dir_all; use utils::id::{NodeId, TenantTimelineId}; use utils::lsn::Lsn; @@ -107,4 +113,59 @@ impl Env { ); Ok(timeline) } + + // This will be dead code when building a non-benchmark target with the + // benchmarking feature enabled. + #[allow(dead_code)] + pub(crate) async fn write_wal( + tli: Arc, + start_lsn: Lsn, + msg_size: usize, + msg_count: usize, + ) -> anyhow::Result { + let (msg_tx, msg_rx) = tokio::sync::mpsc::channel(receive_wal::MSG_QUEUE_SIZE); + let (reply_tx, mut reply_rx) = tokio::sync::mpsc::channel(receive_wal::REPLY_QUEUE_SIZE); + + let end_watch = EndWatch::Commit(tli.get_commit_lsn_watch_rx()); + + WalAcceptor::spawn(tli.wal_residence_guard().await?, msg_rx, reply_tx, Some(0)); + + let prefix = c"p"; + let prefixlen = prefix.to_bytes_with_nul().len(); + assert!(msg_size >= prefixlen); + let message = vec![0; msg_size - prefixlen]; + + let walgen = + &mut WalGenerator::new(LogicalMessageGenerator::new(prefix, &message), start_lsn); + for _ in 0..msg_count { + let (lsn, record) = walgen.next().unwrap(); + + let req = AppendRequest { + h: AppendRequestHeader { + term: 1, + term_start_lsn: start_lsn, + begin_lsn: lsn, + end_lsn: lsn + record.len() as u64, + commit_lsn: lsn, + truncate_lsn: Lsn(0), + proposer_uuid: [0; 16], + }, + wal_data: record, + }; + + let end_lsn = req.h.end_lsn; + + let msg = ProposerAcceptorMessage::AppendRequest(req); + msg_tx.send(msg).await?; + while let Some(reply) = reply_rx.recv().await { + if let AcceptorProposerMessage::AppendResponse(resp) = reply { + if resp.flush_lsn >= end_lsn { + break; + } + } + } + } + + Ok(end_watch) + } } diff --git a/safekeeper/src/timeline.rs b/safekeeper/src/timeline.rs index 2882391074..5eb0bd7146 100644 --- a/safekeeper/src/timeline.rs +++ b/safekeeper/src/timeline.rs @@ -35,7 +35,7 @@ use crate::control_file; use crate::rate_limit::RateLimiter; use crate::receive_wal::WalReceivers; use crate::safekeeper::{AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, TermLsn}; -use crate::send_wal::WalSenders; +use crate::send_wal::{WalSenders, WalSendersTimelineMetricValues}; use crate::state::{EvictionState, TimelineMemState, TimelinePersistentState, TimelineState}; use crate::timeline_guard::ResidenceGuard; use crate::timeline_manager::{AtomicStatus, ManagerCtl}; @@ -712,16 +712,22 @@ impl Timeline { return None; } - let (ps_feedback_count, last_ps_feedback) = self.walsenders.get_ps_feedback_stats(); + let WalSendersTimelineMetricValues { + ps_feedback_counter, + last_ps_feedback, + interpreted_wal_reader_tasks, + } = self.walsenders.info_for_metrics(); + let state = self.read_shared_state().await; Some(FullTimelineInfo { ttid: self.ttid, - ps_feedback_count, + ps_feedback_count: ps_feedback_counter, last_ps_feedback, wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed), timeline_is_active: self.broker_active.load(Ordering::Relaxed), num_computes: self.walreceivers.get_num() as u32, last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed), + interpreted_wal_reader_tasks, epoch_start_lsn: state.sk.term_start_lsn(), mem_state: state.sk.state().inmem.clone(), persisted_state: TimelinePersistentState::clone(state.sk.state()), @@ -740,7 +746,7 @@ impl Timeline { debug_dump::Memory { is_cancelled: self.is_cancelled(), peers_info_len: state.peers_info.0.len(), - walsenders: self.walsenders.get_all(), + walsenders: self.walsenders.get_all_public(), wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed), active: self.broker_active.load(Ordering::Relaxed), num_computes: self.walreceivers.get_num() as u32, diff --git a/safekeeper/src/wal_reader_stream.rs b/safekeeper/src/wal_reader_stream.rs index aea628c208..adac6067da 100644 --- a/safekeeper/src/wal_reader_stream.rs +++ b/safekeeper/src/wal_reader_stream.rs @@ -1,34 +1,16 @@ -use std::sync::Arc; - -use async_stream::try_stream; -use bytes::Bytes; -use futures::Stream; -use postgres_backend::CopyStreamHandlerEnd; -use safekeeper_api::Term; -use std::time::Duration; -use tokio::time::timeout; -use utils::lsn::Lsn; - -use crate::{ - send_wal::{EndWatch, WalSenderGuard}, - timeline::WalResidentTimeline, +use std::{ + pin::Pin, + task::{Context, Poll}, }; -pub(crate) struct WalReaderStreamBuilder { - pub(crate) tli: WalResidentTimeline, - pub(crate) start_pos: Lsn, - pub(crate) end_pos: Lsn, - pub(crate) term: Option, - pub(crate) end_watch: EndWatch, - pub(crate) wal_sender_guard: Arc, -} +use bytes::Bytes; +use futures::{stream::BoxStream, Stream, StreamExt}; +use utils::lsn::Lsn; -impl WalReaderStreamBuilder { - pub(crate) fn start_pos(&self) -> Lsn { - self.start_pos - } -} +use crate::{send_wal::EndWatch, timeline::WalResidentTimeline, wal_storage::WalReader}; +use safekeeper_api::Term; +#[derive(PartialEq, Eq, Debug)] pub(crate) struct WalBytes { /// Raw PG WAL pub(crate) wal: Bytes, @@ -44,106 +26,270 @@ pub(crate) struct WalBytes { pub(crate) available_wal_end_lsn: Lsn, } -impl WalReaderStreamBuilder { - /// Builds a stream of Postgres WAL starting from [`Self::start_pos`]. - /// The stream terminates when the receiver (pageserver) is fully caught up - /// and there's no active computes. - pub(crate) async fn build( - self, - buffer_size: usize, - ) -> anyhow::Result>> { - // TODO(vlad): The code below duplicates functionality from [`crate::send_wal`]. - // We can make the raw WAL sender use this stream too and remove the duplication. - let Self { - tli, - mut start_pos, - mut end_pos, - term, - mut end_watch, - wal_sender_guard, - } = self; - let mut wal_reader = tli.get_walreader(start_pos).await?; - let mut buffer = vec![0; buffer_size]; +struct PositionedWalReader { + start: Lsn, + end: Lsn, + reader: Option, +} - const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1); +/// A streaming WAL reader wrapper which can be reset while running +pub(crate) struct StreamingWalReader { + stream: BoxStream<'static, WalOrReset>, + start_changed_tx: tokio::sync::watch::Sender, +} - Ok(try_stream! { - loop { - let have_something_to_send = end_pos > start_pos; +pub(crate) enum WalOrReset { + Wal(anyhow::Result), + Reset(Lsn), +} - if !have_something_to_send { - // wait for lsn - let res = timeout(POLL_STATE_TIMEOUT, end_watch.wait_for_lsn(start_pos, term)).await; - match res { - Ok(ok) => { - end_pos = ok?; - }, - Err(_) => { - if let EndWatch::Commit(_) = end_watch { - if let Some(remote_consistent_lsn) = wal_sender_guard - .walsenders() - .get_ws_remote_consistent_lsn(wal_sender_guard.id()) - { - if tli.should_walsender_stop(remote_consistent_lsn).await { - // Stop streaming if the receivers are caught up and - // there's no active compute. This causes the loop in - // [`crate::send_interpreted_wal::InterpretedWalSender::run`] - // to exit and terminate the WAL stream. - return; - } - } - } - - continue; - } - } - } - - - assert!( - end_pos > start_pos, - "nothing to send after waiting for WAL" - ); - - // try to send as much as available, capped by the buffer size - let mut chunk_end_pos = start_pos + buffer_size as u64; - // if we went behind available WAL, back off - if chunk_end_pos >= end_pos { - chunk_end_pos = end_pos; - } else { - // If sending not up to end pos, round down to page boundary to - // avoid breaking WAL record not at page boundary, as protocol - // demands. See walsender.c (XLogSendPhysical). - chunk_end_pos = chunk_end_pos - .checked_sub(chunk_end_pos.block_offset()) - .unwrap(); - } - let send_size = (chunk_end_pos.0 - start_pos.0) as usize; - let buffer = &mut buffer[..send_size]; - let send_size: usize; - { - // If uncommitted part is being pulled, check that the term is - // still the expected one. - let _term_guard = if let Some(t) = term { - Some(tli.acquire_term(t).await?) - } else { - None - }; - // Read WAL into buffer. send_size can be additionally capped to - // segment boundary here. - send_size = wal_reader.read(buffer).await? - }; - let wal = Bytes::copy_from_slice(&buffer[..send_size]); - - yield WalBytes { - wal, - wal_start_lsn: start_pos, - wal_end_lsn: start_pos + send_size as u64, - available_wal_end_lsn: end_pos - }; - - start_pos += send_size as u64; - } - }) +impl WalOrReset { + pub(crate) fn get_wal(self) -> Option> { + match self { + WalOrReset::Wal(wal) => Some(wal), + WalOrReset::Reset(_) => None, + } + } +} + +impl StreamingWalReader { + pub(crate) fn new( + tli: WalResidentTimeline, + term: Option, + start: Lsn, + end: Lsn, + end_watch: EndWatch, + buffer_size: usize, + ) -> Self { + let (start_changed_tx, start_changed_rx) = tokio::sync::watch::channel(start); + + let state = WalReaderStreamState { + tli, + wal_reader: PositionedWalReader { + start, + end, + reader: None, + }, + term, + end_watch, + buffer: vec![0; buffer_size], + buffer_size, + }; + + // When a change notification is received while polling the internal + // reader, stop polling the read future and service the change. + let stream = futures::stream::unfold( + (state, start_changed_rx), + |(mut state, mut rx)| async move { + let wal_or_reset = tokio::select! { + read_res = state.read() => { WalOrReset::Wal(read_res) }, + changed_res = rx.changed() => { + if changed_res.is_err() { + return None; + } + + let new_start_pos = rx.borrow_and_update(); + WalOrReset::Reset(*new_start_pos) + } + }; + + if let WalOrReset::Reset(lsn) = wal_or_reset { + state.wal_reader.start = lsn; + state.wal_reader.reader = None; + } + + Some((wal_or_reset, (state, rx))) + }, + ) + .boxed(); + + Self { + stream, + start_changed_tx, + } + } + + /// Reset the stream to a given position. + pub(crate) async fn reset(&mut self, start: Lsn) { + self.start_changed_tx.send(start).unwrap(); + while let Some(wal_or_reset) = self.stream.next().await { + match wal_or_reset { + WalOrReset::Reset(at) => { + // Stream confirmed the reset. + // There may only one ongoing reset at any given time, + // hence the assertion. + assert_eq!(at, start); + break; + } + WalOrReset::Wal(_) => { + // Ignore wal generated before reset was handled + } + } + } + } +} + +impl Stream for StreamingWalReader { + type Item = WalOrReset; + + fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll> { + Pin::new(&mut self.stream).poll_next(cx) + } +} + +struct WalReaderStreamState { + tli: WalResidentTimeline, + wal_reader: PositionedWalReader, + term: Option, + end_watch: EndWatch, + buffer: Vec, + buffer_size: usize, +} + +impl WalReaderStreamState { + async fn read(&mut self) -> anyhow::Result { + // Create reader if needed + if self.wal_reader.reader.is_none() { + self.wal_reader.reader = Some(self.tli.get_walreader(self.wal_reader.start).await?); + } + + let have_something_to_send = self.wal_reader.end > self.wal_reader.start; + if !have_something_to_send { + tracing::debug!( + "Waiting for wal: start={}, end={}", + self.wal_reader.end, + self.wal_reader.start + ); + self.wal_reader.end = self + .end_watch + .wait_for_lsn(self.wal_reader.start, self.term) + .await?; + tracing::debug!( + "Done waiting for wal: start={}, end={}", + self.wal_reader.end, + self.wal_reader.start + ); + } + + assert!( + self.wal_reader.end > self.wal_reader.start, + "nothing to send after waiting for WAL" + ); + + // Calculate chunk size + let mut chunk_end_pos = self.wal_reader.start + self.buffer_size as u64; + if chunk_end_pos >= self.wal_reader.end { + chunk_end_pos = self.wal_reader.end; + } else { + chunk_end_pos = chunk_end_pos + .checked_sub(chunk_end_pos.block_offset()) + .unwrap(); + } + + let send_size = (chunk_end_pos.0 - self.wal_reader.start.0) as usize; + let buffer = &mut self.buffer[..send_size]; + + // Read WAL + let send_size = { + let _term_guard = if let Some(t) = self.term { + Some(self.tli.acquire_term(t).await?) + } else { + None + }; + self.wal_reader + .reader + .as_mut() + .unwrap() + .read(buffer) + .await? + }; + + let wal = Bytes::copy_from_slice(&buffer[..send_size]); + let result = WalBytes { + wal, + wal_start_lsn: self.wal_reader.start, + wal_end_lsn: self.wal_reader.start + send_size as u64, + available_wal_end_lsn: self.wal_reader.end, + }; + + self.wal_reader.start += send_size as u64; + + Ok(result) + } +} + +#[cfg(test)] +mod tests { + use std::str::FromStr; + + use futures::StreamExt; + use postgres_ffi::MAX_SEND_SIZE; + use utils::{ + id::{NodeId, TenantTimelineId}, + lsn::Lsn, + }; + + use crate::{test_utils::Env, wal_reader_stream::StreamingWalReader}; + + #[tokio::test] + async fn test_streaming_wal_reader_reset() { + let _ = env_logger::builder().is_test(true).try_init(); + + const SIZE: usize = 8 * 1024; + const MSG_COUNT: usize = 200; + + let start_lsn = Lsn::from_str("0/149FD18").unwrap(); + let env = Env::new(true).unwrap(); + let tli = env + .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn) + .await + .unwrap(); + + let resident_tli = tli.wal_residence_guard().await.unwrap(); + let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT) + .await + .unwrap(); + let end_pos = end_watch.get(); + + tracing::info!("Doing first round of reads ..."); + + let mut streaming_wal_reader = StreamingWalReader::new( + resident_tli, + None, + start_lsn, + end_pos, + end_watch, + MAX_SEND_SIZE, + ); + + let mut before_reset = Vec::new(); + while let Some(wor) = streaming_wal_reader.next().await { + let wal = wor.get_wal().unwrap().unwrap(); + let stop = wal.available_wal_end_lsn == wal.wal_end_lsn; + before_reset.push(wal); + + if stop { + break; + } + } + + tracing::info!("Resetting the WAL stream ..."); + + streaming_wal_reader.reset(start_lsn).await; + + tracing::info!("Doing second round of reads ..."); + + let mut after_reset = Vec::new(); + while let Some(wor) = streaming_wal_reader.next().await { + let wal = wor.get_wal().unwrap().unwrap(); + let stop = wal.available_wal_end_lsn == wal.wal_end_lsn; + after_reset.push(wal); + + if stop { + break; + } + } + + assert_eq!(before_reset, after_reset); } } diff --git a/safekeeper/tests/walproposer_sim/safekeeper.rs b/safekeeper/tests/walproposer_sim/safekeeper.rs index a99de71a04..e0d593851e 100644 --- a/safekeeper/tests/walproposer_sim/safekeeper.rs +++ b/safekeeper/tests/walproposer_sim/safekeeper.rs @@ -178,6 +178,8 @@ pub fn run_server(os: NodeOs, disk: Arc) -> Result<()> { control_file_save_interval: Duration::from_secs(1), partial_backup_concurrency: 1, eviction_min_resident: Duration::ZERO, + wal_reader_fanout: false, + max_delta_for_fanout: None, }; let mut global = GlobalMap::new(disk, conf.clone())?; From 3d41069dc4002a0444c3cf6afaa87a05d95bcdb5 Mon Sep 17 00:00:00 2001 From: Tristan Partin Date: Wed, 15 Jan 2025 10:26:58 -0600 Subject: [PATCH 18/31] Update pgrx in extension builds to 0.12.9 (#10372) Signed-off-by: Tristan Partin --- compute/compute-node.Dockerfile | 25 +++++++++++++++---------- 1 file changed, 15 insertions(+), 10 deletions(-) diff --git a/compute/compute-node.Dockerfile b/compute/compute-node.Dockerfile index 299f4444a3..1ee159e5df 100644 --- a/compute/compute-node.Dockerfile +++ b/compute/compute-node.Dockerfile @@ -871,7 +871,7 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux chmod +x rustup-init && \ ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \ rm rustup-init && \ - cargo install --locked --version 0.12.6 cargo-pgrx && \ + cargo install --locked --version 0.12.9 cargo-pgrx && \ /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config' USER root @@ -908,19 +908,19 @@ RUN apt update && apt install --no-install-recommends --no-install-suggests -y p mkdir pgrag-src && cd pgrag-src && tar xzf ../pgrag.tar.gz --strip-components=1 -C . && \ \ cd exts/rag && \ - sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ cargo pgrx install --release && \ echo "trusted = true" >> /usr/local/pgsql/share/extension/rag.control && \ \ cd ../rag_bge_small_en_v15 && \ - sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \ REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/bge_small_en_v15.onnx \ cargo pgrx install --release --features remote_onnx && \ echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_bge_small_en_v15.control && \ \ cd ../rag_jina_reranker_v1_tiny_en && \ - sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \ REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/jina_reranker_v1_tiny_en.onnx \ cargo pgrx install --release --features remote_onnx && \ @@ -945,7 +945,8 @@ RUN wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v0.3.3.tar. # against postgres forks that decided to change their ABI name (like us). # With that we can build extensions without forking them and using stock # pgx. As this feature is new few manual version bumps were required. - sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx-tests = "0.12.6"/pgrx-tests = "0.12.9"/g' Cargo.toml && \ cargo pgrx install --release && \ echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_jsonschema.control @@ -963,7 +964,8 @@ ARG PG_VERSION RUN wget https://github.com/supabase/pg_graphql/archive/refs/tags/v1.5.9.tar.gz -O pg_graphql.tar.gz && \ echo "cf768385a41278be1333472204fc0328118644ae443182cf52f7b9b23277e497 pg_graphql.tar.gz" | sha256sum --check && \ mkdir pg_graphql-src && cd pg_graphql-src && tar xzf ../pg_graphql.tar.gz --strip-components=1 -C . && \ - sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "=0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx-tests = "=0.12.6"/pgrx-tests = "=0.12.9"/g' Cargo.toml && \ cargo pgrx install --release && \ # it's needed to enable extension because it uses untrusted C language sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_graphql.control && \ @@ -984,9 +986,8 @@ ARG PG_VERSION RUN wget https://github.com/kelvich/pg_tiktoken/archive/9118dd4549b7d8c0bbc98e04322499f7bf2fa6f7.tar.gz -O pg_tiktoken.tar.gz && \ echo "a5bc447e7920ee149d3c064b8b9f0086c0e83939499753178f7d35788416f628 pg_tiktoken.tar.gz" | sha256sum --check && \ mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \ - # TODO update pgrx version in the pg_tiktoken repo and remove this line - sed -i 's/pgrx = { version = "=0.10.2",/pgrx = { version = "0.11.3",/g' Cargo.toml && \ - sed -i 's/pgrx-tests = "=0.10.2"/pgrx-tests = "0.11.3"/g' Cargo.toml && \ + sed -i 's/pgrx = { version = "=0.12.6",/pgrx = { version = "0.12.9",/g' Cargo.toml && \ + sed -i 's/pgrx-tests = "=0.12.6"/pgrx-tests = "0.12.9"/g' Cargo.toml && \ cargo pgrx install --release && \ echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_tiktoken.control @@ -1028,7 +1029,11 @@ ARG PG_VERSION RUN wget https://github.com/neondatabase/pg_session_jwt/archive/refs/tags/v0.2.0.tar.gz -O pg_session_jwt.tar.gz && \ echo "5ace028e591f2e000ca10afa5b1ca62203ebff014c2907c0ec3b29c36f28a1bb pg_session_jwt.tar.gz" | sha256sum --check && \ mkdir pg_session_jwt-src && cd pg_session_jwt-src && tar xzf ../pg_session_jwt.tar.gz --strip-components=1 -C . && \ - sed -i 's/pgrx = "0.12.6"/pgrx = { version = "=0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \ + sed -i 's/version = "0.12.6"/version = "0.12.9"/g' pgrx-tests/Cargo.toml && \ + sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "=0.12.9", features = [ "unsafe-postgres" ] }/g' pgrx-tests/Cargo.toml && \ + sed -i 's/pgrx-macros = "=0.12.6"/pgrx-macros = "=0.12.9"/g' pgrx-tests/Cargo.toml && \ + sed -i 's/pgrx-pg-config = "=0.12.6"/pgrx-pg-config = "=0.12.9"/g' pgrx-tests/Cargo.toml && \ cargo pgrx install --release ######################################################################################### From efaec6cdf81f71ded5428aeccc6e9f8f607606e0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Arpad=20M=C3=BCller?= Date: Wed, 15 Jan 2025 19:15:30 +0100 Subject: [PATCH 19/31] Add endpoint and storcon cli cmd to set sk scheduling policy (#10400) Implementing the last missing endpoint of #9981, this adds support to set the scheduling policy of an individual safekeeper, as specified in the RFC. However, unlike in the RFC we call the endpoint `scheduling_policy` not `status` Closes #9981. As for why not use the upsert endpoint for this: we want to have the safekeeper upsert endpoint be used for testing and for deploying new safekeepers, but not for changes of the scheduling policy. We don't want to change any of the other fields when marking a safekeeper as decommissioned for example, so we'd have to first fetch them only to then specify them again. Of course one can also design an endpoint where one can omit any field and it doesn't get modified, but it's still not great for observability to put everything into one big "change something about this safekeeper" endpoint. --- control_plane/storcon_cli/src/main.rs | 47 ++++++++++++++++++- libs/pageserver_api/src/controller_api.rs | 7 ++- storage_controller/src/http.rs | 44 ++++++++++++++++- storage_controller/src/persistence.rs | 31 ++++++++++++ storage_controller/src/service.rs | 12 ++++- test_runner/fixtures/neon_fixtures.py | 12 ++++- .../regress/test_storage_controller.py | 11 +++++ 7 files changed, 155 insertions(+), 9 deletions(-) diff --git a/control_plane/storcon_cli/src/main.rs b/control_plane/storcon_cli/src/main.rs index 2ba8f63678..96bfad4c86 100644 --- a/control_plane/storcon_cli/src/main.rs +++ b/control_plane/storcon_cli/src/main.rs @@ -9,8 +9,9 @@ use clap::{Parser, Subcommand}; use pageserver_api::{ controller_api::{ AvailabilityZone, NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse, - SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest, - TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest, + SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest, ShardSchedulingPolicy, + ShardsPreferredAzsRequest, SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, + TenantPolicyRequest, }, models::{ EvictionPolicy, EvictionPolicyLayerAccessThreshold, LocationConfigSecondary, @@ -231,6 +232,13 @@ enum Command { }, /// List safekeepers known to the storage controller Safekeepers {}, + /// Set the scheduling policy of the specified safekeeper + SafekeeperScheduling { + #[arg(long)] + node_id: NodeId, + #[arg(long)] + scheduling_policy: SkSchedulingPolicyArg, + }, } #[derive(Parser)] @@ -283,6 +291,24 @@ impl FromStr for PlacementPolicyArg { } } +#[derive(Debug, Clone)] +struct SkSchedulingPolicyArg(SkSchedulingPolicy); + +impl FromStr for SkSchedulingPolicyArg { + type Err = anyhow::Error; + + fn from_str(s: &str) -> Result { + match s { + "active" => Ok(Self(SkSchedulingPolicy::Active)), + "disabled" => Ok(Self(SkSchedulingPolicy::Disabled)), + "decomissioned" => Ok(Self(SkSchedulingPolicy::Decomissioned)), + _ => Err(anyhow::anyhow!( + "Unknown scheduling policy '{s}', try active,disabled,decomissioned" + )), + } + } +} + #[derive(Debug, Clone)] struct ShardSchedulingPolicyArg(ShardSchedulingPolicy); @@ -1202,6 +1228,23 @@ async fn main() -> anyhow::Result<()> { } println!("{table}"); } + Command::SafekeeperScheduling { + node_id, + scheduling_policy, + } => { + let scheduling_policy = scheduling_policy.0; + storcon_client + .dispatch::( + Method::POST, + format!("control/v1/safekeeper/{node_id}/scheduling_policy"), + Some(SafekeeperSchedulingPolicyRequest { scheduling_policy }), + ) + .await?; + println!( + "Scheduling policy of {node_id} set to {}", + String::from(scheduling_policy) + ); + } } Ok(()) diff --git a/libs/pageserver_api/src/controller_api.rs b/libs/pageserver_api/src/controller_api.rs index f3880cb766..08d1fa55b9 100644 --- a/libs/pageserver_api/src/controller_api.rs +++ b/libs/pageserver_api/src/controller_api.rs @@ -416,8 +416,6 @@ pub struct MetadataHealthListOutdatedResponse { } /// Publicly exposed safekeeper description -/// -/// The `active` flag which we have in the DB is not included on purpose: it is deprecated. #[derive(Serialize, Deserialize, Clone)] pub struct SafekeeperDescribeResponse { pub id: NodeId, @@ -433,6 +431,11 @@ pub struct SafekeeperDescribeResponse { pub scheduling_policy: SkSchedulingPolicy, } +#[derive(Serialize, Deserialize, Clone)] +pub struct SafekeeperSchedulingPolicyRequest { + pub scheduling_policy: SkSchedulingPolicy, +} + #[cfg(test)] mod test { use super::*; diff --git a/storage_controller/src/http.rs b/storage_controller/src/http.rs index 03d8f11992..ac890b008f 100644 --- a/storage_controller/src/http.rs +++ b/storage_controller/src/http.rs @@ -15,7 +15,7 @@ use metrics::{BuildInfo, NeonMetrics}; use pageserver_api::controller_api::{ MetadataHealthListOutdatedRequest, MetadataHealthListOutdatedResponse, MetadataHealthListUnhealthyResponse, MetadataHealthUpdateRequest, MetadataHealthUpdateResponse, - ShardsPreferredAzsRequest, TenantCreateRequest, + SafekeeperSchedulingPolicyRequest, ShardsPreferredAzsRequest, TenantCreateRequest, }; use pageserver_api::models::{ TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest, @@ -1305,6 +1305,35 @@ async fn handle_upsert_safekeeper(mut req: Request) -> Result, +) -> Result, ApiError> { + check_permissions(&req, Scope::Admin)?; + + let body = json_request::(&mut req).await?; + let id = parse_request_param::(&req, "id")?; + + let req = match maybe_forward(req).await { + ForwardOutcome::Forwarded(res) => { + return res; + } + ForwardOutcome::NotForwarded(req) => req, + }; + + let state = get_state(&req); + + state + .service + .set_safekeeper_scheduling_policy(id, body.scheduling_policy) + .await?; + + Ok(Response::builder() + .status(StatusCode::NO_CONTENT) + .body(Body::empty()) + .unwrap()) +} + /// Common wrapper for request handlers that call into Service and will operate on tenants: they must only /// be allowed to run if Service has finished its initial reconciliation. async fn tenant_service_handler( @@ -1873,7 +1902,18 @@ pub fn make_router( }) .post("/control/v1/safekeeper/:id", |r| { // id is in the body - named_request_span(r, handle_upsert_safekeeper, RequestName("v1_safekeeper")) + named_request_span( + r, + handle_upsert_safekeeper, + RequestName("v1_safekeeper_post"), + ) + }) + .post("/control/v1/safekeeper/:id/scheduling_policy", |r| { + named_request_span( + r, + handle_safekeeper_scheduling_policy, + RequestName("v1_safekeeper_status"), + ) }) // Tenant Shard operations .put("/control/v1/tenant/:tenant_shard_id/migrate", |r| { diff --git a/storage_controller/src/persistence.rs b/storage_controller/src/persistence.rs index eb0bfc879e..37bfaf1139 100644 --- a/storage_controller/src/persistence.rs +++ b/storage_controller/src/persistence.rs @@ -1104,6 +1104,37 @@ impl Persistence { }) .await } + + pub(crate) async fn set_safekeeper_scheduling_policy( + &self, + id_: i64, + scheduling_policy_: SkSchedulingPolicy, + ) -> Result<(), DatabaseError> { + use crate::schema::safekeepers::dsl::*; + + self.with_conn(move |conn| -> DatabaseResult<()> { + #[derive(Insertable, AsChangeset)] + #[diesel(table_name = crate::schema::safekeepers)] + struct UpdateSkSchedulingPolicy<'a> { + id: i64, + scheduling_policy: &'a str, + } + let scheduling_policy_ = String::from(scheduling_policy_); + + let rows_affected = diesel::update(safekeepers.filter(id.eq(id_))) + .set(scheduling_policy.eq(scheduling_policy_)) + .execute(conn)?; + + if rows_affected != 1 { + return Err(DatabaseError::Logical(format!( + "unexpected number of rows ({rows_affected})", + ))); + } + + Ok(()) + }) + .await + } } /// Parts of [`crate::tenant_shard::TenantShard`] that are stored durably diff --git a/storage_controller/src/service.rs b/storage_controller/src/service.rs index 57f4cc8463..1d85839881 100644 --- a/storage_controller/src/service.rs +++ b/storage_controller/src/service.rs @@ -47,7 +47,7 @@ use pageserver_api::{ AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability, NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy, SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest, - ShardsPreferredAzsResponse, TenantCreateRequest, TenantCreateResponse, + ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest, TenantShardMigrateResponse, @@ -7651,6 +7651,16 @@ impl Service { self.persistence.safekeeper_upsert(record).await } + pub(crate) async fn set_safekeeper_scheduling_policy( + &self, + id: i64, + scheduling_policy: SkSchedulingPolicy, + ) -> Result<(), DatabaseError> { + self.persistence + .set_safekeeper_scheduling_policy(id, scheduling_policy) + .await + } + pub(crate) async fn update_shards_preferred_azs( &self, req: ShardsPreferredAzsRequest, diff --git a/test_runner/fixtures/neon_fixtures.py b/test_runner/fixtures/neon_fixtures.py index c47739cd81..c3950e9bf7 100644 --- a/test_runner/fixtures/neon_fixtures.py +++ b/test_runner/fixtures/neon_fixtures.py @@ -2336,6 +2336,14 @@ class NeonStorageController(MetricsGetter, LogUtils): json=body, ) + def safekeeper_scheduling_policy(self, id: int, scheduling_policy: str): + self.request( + "POST", + f"{self.api}/control/v1/safekeeper/{id}/scheduling_policy", + headers=self.headers(TokenScope.ADMIN), + json={"id": id, "scheduling_policy": scheduling_policy}, + ) + def get_safekeeper(self, id: int) -> dict[str, Any] | None: try: response = self.request( @@ -4135,7 +4143,7 @@ class Endpoint(PgProtocol, LogUtils): # Checkpoints running endpoint and returns pg_wal size in MB. def get_pg_wal_size(self): - log.info(f'checkpointing at LSN {self.safe_psql("select pg_current_wal_lsn()")[0][0]}') + log.info(f"checkpointing at LSN {self.safe_psql('select pg_current_wal_lsn()')[0][0]}") self.safe_psql("checkpoint") assert self.pgdata_dir is not None # please mypy return get_dir_size(self.pgdata_dir / "pg_wal") / 1024 / 1024 @@ -4975,7 +4983,7 @@ def logical_replication_sync( if res: log.info(f"subscriber_lsn={res}") subscriber_lsn = Lsn(res) - log.info(f"Subscriber LSN={subscriber_lsn}, publisher LSN={ publisher_lsn}") + log.info(f"Subscriber LSN={subscriber_lsn}, publisher LSN={publisher_lsn}") if subscriber_lsn >= publisher_lsn: return subscriber_lsn time.sleep(0.5) diff --git a/test_runner/regress/test_storage_controller.py b/test_runner/regress/test_storage_controller.py index b5d109559f..b1e1fd81d6 100644 --- a/test_runner/regress/test_storage_controller.py +++ b/test_runner/regress/test_storage_controller.py @@ -3208,6 +3208,17 @@ def test_safekeeper_deployment_time_update(neon_env_builder: NeonEnvBuilder): assert eq_safekeeper_records(body, inserted_now) + # some small tests for the scheduling policy querying and returning APIs + newest_info = target.get_safekeeper(inserted["id"]) + assert newest_info + assert newest_info["scheduling_policy"] == "Disabled" + target.safekeeper_scheduling_policy(inserted["id"], "Decomissioned") + newest_info = target.get_safekeeper(inserted["id"]) + assert newest_info + assert newest_info["scheduling_policy"] == "Decomissioned" + # Ensure idempotency + target.safekeeper_scheduling_policy(inserted["id"], "Decomissioned") + def eq_safekeeper_records(a: dict[str, Any], b: dict[str, Any]) -> bool: compared = [dict(a), dict(b)] From fb0e2acb2f8333d279371da84c17e3e6db5c31f2 Mon Sep 17 00:00:00 2001 From: John Spray Date: Wed, 15 Jan 2025 19:07:22 +0000 Subject: [PATCH 20/31] pageserver: add `page_trace` API for debugging (#10293) ## Problem When a pageserver is receiving high rates of requests, we don't have a good way to efficiently discover what the client's access pattern is. Closes: https://github.com/neondatabase/neon/issues/10275 ## Summary of changes - Add `/v1/tenant/x/timeline/y/page_trace?size_limit_bytes=...&time_limit_secs=...` API, which returns a binary buffer. - Add `pagectl page-trace` tool to decode and analyze the output. --------- Co-authored-by: Erik Grinaker --- Cargo.lock | 3 ++ libs/pageserver_api/src/key.rs | 4 +- libs/pageserver_api/src/models.rs | 19 +++++++- pageserver/Cargo.toml | 1 + pageserver/ctl/Cargo.toml | 2 + pageserver/ctl/src/main.rs | 4 ++ pageserver/ctl/src/page_trace.rs | 73 +++++++++++++++++++++++++++++++ pageserver/src/http/routes.rs | 72 ++++++++++++++++++++++++++++++ pageserver/src/page_service.rs | 15 +++++++ pageserver/src/tenant/timeline.rs | 9 +++- 10 files changed, 199 insertions(+), 3 deletions(-) create mode 100644 pageserver/ctl/src/page_trace.rs diff --git a/Cargo.lock b/Cargo.lock index afe16ff848..3f184ebe0b 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -3981,9 +3981,11 @@ name = "pagectl" version = "0.1.0" dependencies = [ "anyhow", + "bincode", "camino", "clap", "humantime", + "itertools 0.10.5", "pageserver", "pageserver_api", "postgres_ffi", @@ -4005,6 +4007,7 @@ dependencies = [ "arc-swap", "async-compression", "async-stream", + "bincode", "bit_field", "byteorder", "bytes", diff --git a/libs/pageserver_api/src/key.rs b/libs/pageserver_api/src/key.rs index 328dea5dec..dbd45da314 100644 --- a/libs/pageserver_api/src/key.rs +++ b/libs/pageserver_api/src/key.rs @@ -24,7 +24,9 @@ pub struct Key { /// When working with large numbers of Keys in-memory, it is more efficient to handle them as i128 than as /// a struct of fields. -#[derive(Clone, Copy, Hash, PartialEq, Eq, Ord, PartialOrd, Serialize, Deserialize, Debug)] +#[derive( + Clone, Copy, Default, Hash, PartialEq, Eq, Ord, PartialOrd, Serialize, Deserialize, Debug, +)] pub struct CompactKey(i128); /// The storage key size. diff --git a/libs/pageserver_api/src/models.rs b/libs/pageserver_api/src/models.rs index 9af6c4021d..87e8df2ab6 100644 --- a/libs/pageserver_api/src/models.rs +++ b/libs/pageserver_api/src/models.rs @@ -29,7 +29,7 @@ use utils::{ }; use crate::{ - key::Key, + key::{CompactKey, Key}, reltag::RelTag, shard::{ShardCount, ShardStripeSize, TenantShardId}, }; @@ -1981,6 +1981,23 @@ impl PagestreamBeMessage { } } +#[derive(Debug, Serialize, Deserialize)] +pub struct PageTraceEvent { + pub key: CompactKey, + pub effective_lsn: Lsn, + pub time: SystemTime, +} + +impl Default for PageTraceEvent { + fn default() -> Self { + Self { + key: Default::default(), + effective_lsn: Default::default(), + time: std::time::UNIX_EPOCH, + } + } +} + #[cfg(test)] mod tests { use serde_json::json; diff --git a/pageserver/Cargo.toml b/pageserver/Cargo.toml index 8547746d94..9195951191 100644 --- a/pageserver/Cargo.toml +++ b/pageserver/Cargo.toml @@ -16,6 +16,7 @@ arc-swap.workspace = true async-compression.workspace = true async-stream.workspace = true bit_field.workspace = true +bincode.workspace = true byteorder.workspace = true bytes.workspace = true camino.workspace = true diff --git a/pageserver/ctl/Cargo.toml b/pageserver/ctl/Cargo.toml index 39ca47568c..7b70f0dc87 100644 --- a/pageserver/ctl/Cargo.toml +++ b/pageserver/ctl/Cargo.toml @@ -8,9 +8,11 @@ license.workspace = true [dependencies] anyhow.workspace = true +bincode.workspace = true camino.workspace = true clap = { workspace = true, features = ["string"] } humantime.workspace = true +itertools.workspace = true pageserver = { path = ".." } pageserver_api.workspace = true remote_storage = { path = "../../libs/remote_storage" } diff --git a/pageserver/ctl/src/main.rs b/pageserver/ctl/src/main.rs index a0aac89dc8..353b4bd2f9 100644 --- a/pageserver/ctl/src/main.rs +++ b/pageserver/ctl/src/main.rs @@ -9,7 +9,9 @@ mod index_part; mod key; mod layer_map_analyzer; mod layers; +mod page_trace; +use page_trace::PageTraceCmd; use std::{ str::FromStr, time::{Duration, SystemTime}, @@ -64,6 +66,7 @@ enum Commands { Layer(LayerCmd), /// Debug print a hex key found from logs Key(key::DescribeKeyCommand), + PageTrace(PageTraceCmd), } /// Read and update pageserver metadata file @@ -183,6 +186,7 @@ async fn main() -> anyhow::Result<()> { .await?; } Commands::Key(dkc) => dkc.execute(), + Commands::PageTrace(cmd) => page_trace::main(&cmd)?, }; Ok(()) } diff --git a/pageserver/ctl/src/page_trace.rs b/pageserver/ctl/src/page_trace.rs new file mode 100644 index 0000000000..da0de72fd9 --- /dev/null +++ b/pageserver/ctl/src/page_trace.rs @@ -0,0 +1,73 @@ +use std::collections::HashMap; +use std::io::BufReader; + +use camino::Utf8PathBuf; +use clap::Parser; +use itertools::Itertools as _; +use pageserver_api::key::{CompactKey, Key}; +use pageserver_api::models::PageTraceEvent; +use pageserver_api::reltag::RelTag; + +/// Parses a page trace (as emitted by the `page_trace` timeline API), and outputs stats. +#[derive(Parser)] +pub(crate) struct PageTraceCmd { + /// Trace input file. + path: Utf8PathBuf, +} + +pub(crate) fn main(cmd: &PageTraceCmd) -> anyhow::Result<()> { + let mut file = BufReader::new(std::fs::OpenOptions::new().read(true).open(&cmd.path)?); + let mut events: Vec = Vec::new(); + loop { + match bincode::deserialize_from(&mut file) { + Ok(event) => events.push(event), + Err(err) => { + if let bincode::ErrorKind::Io(ref err) = *err { + if err.kind() == std::io::ErrorKind::UnexpectedEof { + break; + } + } + return Err(err.into()); + } + } + } + + let mut reads_by_relation: HashMap = HashMap::new(); + let mut reads_by_key: HashMap = HashMap::new(); + + for event in events { + let key = Key::from_compact(event.key); + let reltag = RelTag { + spcnode: key.field2, + dbnode: key.field3, + relnode: key.field4, + forknum: key.field5, + }; + + *reads_by_relation.entry(reltag).or_default() += 1; + *reads_by_key.entry(event.key).or_default() += 1; + } + + let multi_read_keys = reads_by_key + .into_iter() + .filter(|(_, count)| *count > 1) + .sorted_by_key(|(key, count)| (-*count, *key)) + .collect_vec(); + + println!("Multi-read keys: {}", multi_read_keys.len()); + for (key, count) in multi_read_keys { + println!(" {key}: {count}"); + } + + let reads_by_relation = reads_by_relation + .into_iter() + .sorted_by_key(|(rel, count)| (-*count, *rel)) + .collect_vec(); + + println!("Reads by relation:"); + for (reltag, count) in reads_by_relation { + println!(" {reltag}: {count}"); + } + + Ok(()) +} diff --git a/pageserver/src/http/routes.rs b/pageserver/src/http/routes.rs index 94e0b101bd..33b2d04588 100644 --- a/pageserver/src/http/routes.rs +++ b/pageserver/src/http/routes.rs @@ -27,6 +27,7 @@ use pageserver_api::models::LocationConfigMode; use pageserver_api::models::LsnLease; use pageserver_api::models::LsnLeaseRequest; use pageserver_api::models::OffloadedTimelineInfo; +use pageserver_api::models::PageTraceEvent; use pageserver_api::models::ShardParameters; use pageserver_api::models::TenantConfigPatchRequest; use pageserver_api::models::TenantDetails; @@ -51,7 +52,9 @@ use pageserver_api::shard::TenantShardId; use remote_storage::DownloadError; use remote_storage::GenericRemoteStorage; use remote_storage::TimeTravelError; +use scopeguard::defer; use tenant_size_model::{svg::SvgBranchKind, SizeResult, StorageModel}; +use tokio::time::Instant; use tokio_util::io::StreamReader; use tokio_util::sync::CancellationToken; use tracing::*; @@ -1521,6 +1524,71 @@ async fn timeline_gc_unblocking_handler( block_or_unblock_gc(request, false).await } +/// Traces GetPage@LSN requests for a timeline, and emits metadata in an efficient binary encoding. +/// Use the `pagectl page-trace` command to decode and analyze the output. +async fn timeline_page_trace_handler( + request: Request, + cancel: CancellationToken, +) -> Result, ApiError> { + let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?; + let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?; + let state = get_state(&request); + check_permission(&request, None)?; + + let size_limit: usize = parse_query_param(&request, "size_limit_bytes")?.unwrap_or(1024 * 1024); + let time_limit_secs: u64 = parse_query_param(&request, "time_limit_secs")?.unwrap_or(5); + + // Convert size limit to event limit based on the serialized size of an event. The event size is + // fixed, as the default bincode serializer uses fixed-width integer encoding. + let event_size = bincode::serialize(&PageTraceEvent::default()) + .map_err(|err| ApiError::InternalServerError(err.into()))? + .len(); + let event_limit = size_limit / event_size; + + let timeline = + active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id) + .await?; + + // Install a page trace, unless one is already in progress. We just use a buffered channel, + // which may 2x the memory usage in the worst case, but it's still bounded. + let (trace_tx, mut trace_rx) = tokio::sync::mpsc::channel(event_limit); + let cur = timeline.page_trace.load(); + let installed = cur.is_none() + && timeline + .page_trace + .compare_and_swap(cur, Some(Arc::new(trace_tx))) + .is_none(); + if !installed { + return Err(ApiError::Conflict("page trace already active".to_string())); + } + defer!(timeline.page_trace.store(None)); // uninstall on return + + // Collect the trace and return it to the client. We could stream the response, but this is + // simple and fine. + let mut body = Vec::with_capacity(size_limit); + let deadline = Instant::now() + Duration::from_secs(time_limit_secs); + + while body.len() < size_limit { + tokio::select! { + event = trace_rx.recv() => { + let Some(event) = event else { + break; // shouldn't happen (sender doesn't close, unless timeline dropped) + }; + bincode::serialize_into(&mut body, &event) + .map_err(|err| ApiError::InternalServerError(err.into()))?; + } + _ = tokio::time::sleep_until(deadline) => break, // time limit reached + _ = cancel.cancelled() => return Err(ApiError::Cancelled), + } + } + + Ok(Response::builder() + .status(StatusCode::OK) + .header(header::CONTENT_TYPE, "application/octet-stream") + .body(hyper::Body::from(body)) + .unwrap()) +} + /// Adding a block is `POST ../block_gc`, removing a block is `POST ../unblock_gc`. /// /// Both are technically unsafe because they might fire off index uploads, thus they are POST. @@ -3479,6 +3547,10 @@ pub fn make_router( "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/unblock_gc", |r| api_handler(r, timeline_gc_unblocking_handler), ) + .get( + "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/page_trace", + |r| api_handler(r, timeline_page_trace_handler), + ) .post("/v1/tenant/:tenant_shard_id/heatmap_upload", |r| { api_handler(r, secondary_upload_handler) }) diff --git a/pageserver/src/page_service.rs b/pageserver/src/page_service.rs index b3e18fed99..da4180a927 100644 --- a/pageserver/src/page_service.rs +++ b/pageserver/src/page_service.rs @@ -67,6 +67,7 @@ use crate::tenant::PageReconstructError; use crate::tenant::Timeline; use crate::{basebackup, timed_after_cancellation}; use pageserver_api::key::rel_block_to_key; +use pageserver_api::models::PageTraceEvent; use pageserver_api::reltag::SlruKind; use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID; use postgres_ffi::BLCKSZ; @@ -1718,6 +1719,20 @@ impl PageServerHandler { .query_metrics .observe_getpage_batch_start(requests.len()); + // If a page trace is running, submit an event for this request. + if let Some(page_trace) = timeline.page_trace.load().as_ref() { + let time = SystemTime::now(); + for batch in &requests { + let key = rel_block_to_key(batch.req.rel, batch.req.blkno).to_compact(); + // Ignore error (trace buffer may be full or tracer may have disconnected). + _ = page_trace.try_send(PageTraceEvent { + key, + effective_lsn, + time, + }); + } + } + let results = timeline .get_rel_page_at_lsn_batched( requests.iter().map(|p| (&p.req.rel, &p.req.blkno)), diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs index 4aa6b7a05a..d6ae11e67d 100644 --- a/pageserver/src/tenant/timeline.rs +++ b/pageserver/src/tenant/timeline.rs @@ -14,7 +14,7 @@ pub mod uninit; mod walreceiver; use anyhow::{anyhow, bail, ensure, Context, Result}; -use arc_swap::ArcSwap; +use arc_swap::{ArcSwap, ArcSwapOption}; use bytes::Bytes; use camino::Utf8Path; use chrono::{DateTime, Utc}; @@ -23,6 +23,7 @@ use fail::fail_point; use handle::ShardTimelineId; use offload::OffloadError; use once_cell::sync::Lazy; +use pageserver_api::models::PageTraceEvent; use pageserver_api::{ config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD, key::{ @@ -42,6 +43,7 @@ use rand::Rng; use remote_storage::DownloadError; use serde_with::serde_as; use storage_broker::BrokerClientChannel; +use tokio::sync::mpsc::Sender; use tokio::{ runtime::Handle, sync::{oneshot, watch}, @@ -433,6 +435,9 @@ pub struct Timeline { /// Cf. [`crate::tenant::CreateTimelineIdempotency`]. pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency, + + /// If Some, collects GetPage metadata for an ongoing PageTrace. + pub(crate) page_trace: ArcSwapOption>, } pub type TimelineDeleteProgress = Arc>; @@ -2380,6 +2385,8 @@ impl Timeline { attach_wal_lag_cooldown, create_idempotency, + + page_trace: Default::default(), }; result.repartition_threshold = From 55a68b28a27b27810255212994e601050b73917e Mon Sep 17 00:00:00 2001 From: Gleb Novikov Date: Wed, 15 Jan 2025 20:51:09 +0000 Subject: [PATCH 21/31] fast import: restore to neondb (not postgres) database (#10251) ## Problem `postgres` is system database at neon, so we need to do `pg_restore` into `neondb` instead https://github.com/neondatabase/cloud/issues/22100 ## Summary of changes Changed fast_import a little bit: 1. After succesfull connection creating `neondb` in postgres instance 2. Changed restore connstring to use new db 3. Added optional `source_connection_string`, which allows to skip `s3_prefix` and just connect directly. 4. Added `-i` that stops process until sigterm ## TODO - [x] test image in cplane e2e - [ ] Change import job image back to latest after this merged (partial revert of https://github.com/neondatabase/cloud/pull/22338) --- compute_tools/src/bin/fast_import.rs | 227 +++++++++++++++++---------- 1 file changed, 147 insertions(+), 80 deletions(-) diff --git a/compute_tools/src/bin/fast_import.rs b/compute_tools/src/bin/fast_import.rs index f554362751..5b008f8182 100644 --- a/compute_tools/src/bin/fast_import.rs +++ b/compute_tools/src/bin/fast_import.rs @@ -31,7 +31,7 @@ use camino::{Utf8Path, Utf8PathBuf}; use clap::Parser; use compute_tools::extension_server::{get_pg_version, PostgresMajorVersion}; use nix::unistd::Pid; -use tracing::{info, info_span, warn, Instrument}; +use tracing::{error, info, info_span, warn, Instrument}; use utils::fs_ext::is_directory_empty; #[path = "fast_import/aws_s3_sync.rs"] @@ -41,12 +41,19 @@ mod child_stdio_to_log; #[path = "fast_import/s3_uri.rs"] mod s3_uri; +const PG_WAIT_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(600); +const PG_WAIT_RETRY_INTERVAL: std::time::Duration = std::time::Duration::from_millis(300); + #[derive(clap::Parser)] struct Args { #[clap(long)] working_directory: Utf8PathBuf, #[clap(long, env = "NEON_IMPORTER_S3_PREFIX")] - s3_prefix: s3_uri::S3Uri, + s3_prefix: Option, + #[clap(long)] + source_connection_string: Option, + #[clap(short, long)] + interactive: bool, #[clap(long)] pg_bin_dir: Utf8PathBuf, #[clap(long)] @@ -77,30 +84,70 @@ pub(crate) async fn main() -> anyhow::Result<()> { info!("starting"); - let Args { - working_directory, - s3_prefix, - pg_bin_dir, - pg_lib_dir, - } = Args::parse(); + let args = Args::parse(); - let aws_config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await; + // Validate arguments + if args.s3_prefix.is_none() && args.source_connection_string.is_none() { + anyhow::bail!("either s3_prefix or source_connection_string must be specified"); + } + if args.s3_prefix.is_some() && args.source_connection_string.is_some() { + anyhow::bail!("only one of s3_prefix or source_connection_string can be specified"); + } - let spec: Spec = { - let spec_key = s3_prefix.append("/spec.json"); - let s3_client = aws_sdk_s3::Client::new(&aws_config); - let object = s3_client - .get_object() - .bucket(&spec_key.bucket) - .key(spec_key.key) - .send() - .await - .context("get spec from s3")? - .body - .collect() - .await - .context("download spec body")?; - serde_json::from_slice(&object.into_bytes()).context("parse spec as json")? + let working_directory = args.working_directory; + let pg_bin_dir = args.pg_bin_dir; + let pg_lib_dir = args.pg_lib_dir; + + // Initialize AWS clients only if s3_prefix is specified + let (aws_config, kms_client) = if args.s3_prefix.is_some() { + let config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await; + let kms = aws_sdk_kms::Client::new(&config); + (Some(config), Some(kms)) + } else { + (None, None) + }; + + // Get source connection string either from S3 spec or direct argument + let source_connection_string = if let Some(s3_prefix) = &args.s3_prefix { + let spec: Spec = { + let spec_key = s3_prefix.append("/spec.json"); + let s3_client = aws_sdk_s3::Client::new(aws_config.as_ref().unwrap()); + let object = s3_client + .get_object() + .bucket(&spec_key.bucket) + .key(spec_key.key) + .send() + .await + .context("get spec from s3")? + .body + .collect() + .await + .context("download spec body")?; + serde_json::from_slice(&object.into_bytes()).context("parse spec as json")? + }; + + match spec.encryption_secret { + EncryptionSecret::KMS { key_id } => { + let mut output = kms_client + .unwrap() + .decrypt() + .key_id(key_id) + .ciphertext_blob(aws_sdk_s3::primitives::Blob::new( + spec.source_connstring_ciphertext_base64, + )) + .send() + .await + .context("decrypt source connection string")?; + let plaintext = output + .plaintext + .take() + .context("get plaintext source connection string")?; + String::from_utf8(plaintext.into_inner()) + .context("parse source connection string as utf8")? + } + } + } else { + args.source_connection_string.unwrap() }; match tokio::fs::create_dir(&working_directory).await { @@ -123,15 +170,6 @@ pub(crate) async fn main() -> anyhow::Result<()> { .await .context("create pgdata directory")?; - // - // Setup clients - // - let aws_config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await; - let kms_client = aws_sdk_kms::Client::new(&aws_config); - - // - // Initialize pgdata - // let pgbin = pg_bin_dir.join("postgres"); let pg_version = match get_pg_version(pgbin.as_ref()) { PostgresMajorVersion::V14 => 14, @@ -170,7 +208,13 @@ pub(crate) async fn main() -> anyhow::Result<()> { .args(["-c", &format!("max_parallel_workers={nproc}")]) .args(["-c", &format!("max_parallel_workers_per_gather={nproc}")]) .args(["-c", &format!("max_worker_processes={nproc}")]) - .args(["-c", "effective_io_concurrency=100"]) + .args([ + "-c", + &format!( + "effective_io_concurrency={}", + if cfg!(target_os = "macos") { 0 } else { 100 } + ), + ]) .env_clear() .stdout(std::process::Stdio::piped()) .stderr(std::process::Stdio::piped()) @@ -185,44 +229,58 @@ pub(crate) async fn main() -> anyhow::Result<()> { ) .instrument(info_span!("postgres")), ); + + // Create neondb database in the running postgres let restore_pg_connstring = format!("host=localhost port=5432 user={superuser} dbname=postgres"); + + let start_time = std::time::Instant::now(); + loop { - let res = tokio_postgres::connect(&restore_pg_connstring, tokio_postgres::NoTls).await; - if res.is_ok() { - info!("postgres is ready, could connect to it"); - break; + if start_time.elapsed() > PG_WAIT_TIMEOUT { + error!( + "timeout exceeded: failed to poll postgres and create database within 10 minutes" + ); + std::process::exit(1); + } + + match tokio_postgres::connect(&restore_pg_connstring, tokio_postgres::NoTls).await { + Ok((client, connection)) => { + // Spawn the connection handling task to maintain the connection + tokio::spawn(async move { + if let Err(e) = connection.await { + warn!("connection error: {}", e); + } + }); + + match client.simple_query("CREATE DATABASE neondb;").await { + Ok(_) => { + info!("created neondb database"); + break; + } + Err(e) => { + warn!( + "failed to create database: {}, retying in {}s", + e, + PG_WAIT_RETRY_INTERVAL.as_secs_f32() + ); + tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await; + continue; + } + } + } + Err(_) => { + info!( + "postgres not ready yet, retrying in {}s", + PG_WAIT_RETRY_INTERVAL.as_secs_f32() + ); + tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await; + continue; + } } } - // - // Decrypt connection string - // - let source_connection_string = { - match spec.encryption_secret { - EncryptionSecret::KMS { key_id } => { - let mut output = kms_client - .decrypt() - .key_id(key_id) - .ciphertext_blob(aws_sdk_s3::primitives::Blob::new( - spec.source_connstring_ciphertext_base64, - )) - .send() - .await - .context("decrypt source connection string")?; - let plaintext = output - .plaintext - .take() - .context("get plaintext source connection string")?; - String::from_utf8(plaintext.into_inner()) - .context("parse source connection string as utf8")? - } - } - }; - - // - // Start the work - // + let restore_pg_connstring = restore_pg_connstring.replace("dbname=postgres", "dbname=neondb"); let dumpdir = working_directory.join("dumpdir"); @@ -310,6 +368,12 @@ pub(crate) async fn main() -> anyhow::Result<()> { } } + // If interactive mode, wait for Ctrl+C + if args.interactive { + info!("Running in interactive mode. Press Ctrl+C to shut down."); + tokio::signal::ctrl_c().await.context("wait for ctrl-c")?; + } + info!("shutdown postgres"); { nix::sys::signal::kill( @@ -325,21 +389,24 @@ pub(crate) async fn main() -> anyhow::Result<()> { .context("wait for postgres to shut down")?; } - info!("upload pgdata"); - aws_s3_sync::sync(Utf8Path::new(&pgdata_dir), &s3_prefix.append("/pgdata/")) - .await - .context("sync dump directory to destination")?; - - info!("write status"); - { - let status_dir = working_directory.join("status"); - std::fs::create_dir(&status_dir).context("create status directory")?; - let status_file = status_dir.join("pgdata"); - std::fs::write(&status_file, serde_json::json!({"done": true}).to_string()) - .context("write status file")?; - aws_s3_sync::sync(&status_dir, &s3_prefix.append("/status/")) + // Only sync if s3_prefix was specified + if let Some(s3_prefix) = args.s3_prefix { + info!("upload pgdata"); + aws_s3_sync::sync(Utf8Path::new(&pgdata_dir), &s3_prefix.append("/pgdata/")) .await - .context("sync status directory to destination")?; + .context("sync dump directory to destination")?; + + info!("write status"); + { + let status_dir = working_directory.join("status"); + std::fs::create_dir(&status_dir).context("create status directory")?; + let status_file = status_dir.join("pgdata"); + std::fs::write(&status_file, serde_json::json!({"done": true}).to_string()) + .context("write status file")?; + aws_s3_sync::sync(&status_dir, &s3_prefix.append("/status/")) + .await + .context("sync status directory to destination")?; + } } Ok(()) From a753349cb0d8a76b7b5884c216236f6e57357d30 Mon Sep 17 00:00:00 2001 From: "Alex Chi Z." <4198311+skyzh@users.noreply.github.com> Date: Wed, 15 Jan 2025 17:04:06 -0500 Subject: [PATCH 22/31] feat(pageserver): validate data integrity during gc-compaction (#10131) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Problem part of https://github.com/neondatabase/neon/issues/9114 part of investigation of https://github.com/neondatabase/neon/issues/10049 ## Summary of changes * If `cfg!(test) or cfg!(feature = testing)`, then we will always try generating an image to ensure the history is replayable, but not put the image layer into the final layer results, therefore discovering wrong key history before we hit a read error. * I suspect it's easier to trigger some races if gc-compaction is continuously run on a timeline, so I increased the frequency to twice per 10 churns. * Also, create branches in gc-compaction smoke tests to get more test coverage. --------- Signed-off-by: Alex Chi Z Co-authored-by: Arpad Müller --- pageserver/src/tenant/timeline/compaction.rs | 77 +++++++++++++++----- test_runner/fixtures/workload.py | 16 ++++ test_runner/regress/test_compaction.py | 20 ++++- 3 files changed, 89 insertions(+), 24 deletions(-) diff --git a/pageserver/src/tenant/timeline/compaction.rs b/pageserver/src/tenant/timeline/compaction.rs index 05f8d476f9..2042a18e96 100644 --- a/pageserver/src/tenant/timeline/compaction.rs +++ b/pageserver/src/tenant/timeline/compaction.rs @@ -1776,7 +1776,10 @@ impl Timeline { base_img_from_ancestor: Option<(Key, Lsn, Bytes)>, ) -> anyhow::Result { // Pre-checks for the invariants - if cfg!(debug_assertions) { + + let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing"); + + if debug_mode { for (log_key, _, _) in full_history { assert_eq!(log_key, &key, "mismatched key"); } @@ -1922,15 +1925,19 @@ impl Timeline { output } + let mut key_exists = false; for (i, split_for_lsn) in split_history.into_iter().enumerate() { // TODO: there could be image keys inside the splits, and we can compute records_since_last_image accordingly. records_since_last_image += split_for_lsn.len(); - let generate_image = if i == 0 && !has_ancestor { + // Whether to produce an image into the final layer files + let produce_image = if i == 0 && !has_ancestor { // We always generate images for the first batch (below horizon / lowest retain_lsn) true } else if i == batch_cnt - 1 { // Do not generate images for the last batch (above horizon) false + } else if records_since_last_image == 0 { + false } else if records_since_last_image >= delta_threshold_cnt { // Generate images when there are too many records true @@ -1945,29 +1952,45 @@ impl Timeline { break; } } - if let Some((_, _, val)) = replay_history.first() { - if !val.will_init() { - return Err(anyhow::anyhow!("invalid history, no base image")).with_context( - || { - generate_debug_trace( - Some(&replay_history), - full_history, - retain_lsn_below_horizon, - horizon, - ) - }, - ); - } + if replay_history.is_empty() && !key_exists { + // The key does not exist at earlier LSN, we can skip this iteration. + retention.push(Vec::new()); + continue; + } else { + key_exists = true; } - if generate_image && records_since_last_image > 0 { + let Some((_, _, val)) = replay_history.first() else { + unreachable!("replay history should not be empty once it exists") + }; + if !val.will_init() { + return Err(anyhow::anyhow!("invalid history, no base image")).with_context(|| { + generate_debug_trace( + Some(&replay_history), + full_history, + retain_lsn_below_horizon, + horizon, + ) + }); + } + // Whether to reconstruct the image. In debug mode, we will generate an image + // at every retain_lsn to ensure data is not corrupted, but we won't put the + // image into the final layer. + let generate_image = produce_image || debug_mode; + if produce_image { records_since_last_image = 0; - let replay_history_for_debug = if cfg!(debug_assertions) { + } + let img_and_lsn = if generate_image { + let replay_history_for_debug = if debug_mode { Some(replay_history.clone()) } else { None }; let replay_history_for_debug_ref = replay_history_for_debug.as_deref(); - let history = std::mem::take(&mut replay_history); + let history = if produce_image { + std::mem::take(&mut replay_history) + } else { + replay_history.clone() + }; let mut img = None; let mut records = Vec::with_capacity(history.len()); if let (_, lsn, Value::Image(val)) = history.first().as_ref().unwrap() { @@ -2004,8 +2027,20 @@ impl Timeline { } records.reverse(); let state = ValueReconstructState { img, records }; - let request_lsn = lsn_split_points[i]; // last batch does not generate image so i is always in range + // last batch does not generate image so i is always in range, unless we force generate + // an image during testing + let request_lsn = if i >= lsn_split_points.len() { + Lsn::MAX + } else { + lsn_split_points[i] + }; let img = self.reconstruct_value(key, request_lsn, state).await?; + Some((request_lsn, img)) + } else { + None + }; + if produce_image { + let (request_lsn, img) = img_and_lsn.unwrap(); replay_history.push((key, request_lsn, Value::Image(img.clone()))); retention.push(vec![(request_lsn, Value::Image(img))]); } else { @@ -2273,6 +2308,8 @@ impl Timeline { let compact_key_range = job.compact_key_range; let compact_lsn_range = job.compact_lsn_range; + let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing"); + info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compact_key_range={}..{}, compact_lsn_range={}..{}", compact_key_range.start, compact_key_range.end, compact_lsn_range.start, compact_lsn_range.end); scopeguard::defer! { @@ -2398,7 +2435,7 @@ impl Timeline { .first() .copied() .unwrap_or(job_desc.gc_cutoff); - if cfg!(debug_assertions) { + if debug_mode { assert_eq!( res, job_desc diff --git a/test_runner/fixtures/workload.py b/test_runner/fixtures/workload.py index 1b8c9fef44..eea0ec2b95 100644 --- a/test_runner/fixtures/workload.py +++ b/test_runner/fixtures/workload.py @@ -53,6 +53,22 @@ class Workload: self._endpoint: Endpoint | None = None self._endpoint_opts = endpoint_opts or {} + def branch( + self, + timeline_id: TimelineId, + branch_name: str | None = None, + endpoint_opts: dict[str, Any] | None = None, + ) -> Workload: + """ + Checkpoint the current status of the workload in case of branching + """ + branch_workload = Workload( + self.env, self.tenant_id, timeline_id, branch_name, endpoint_opts + ) + branch_workload.expect_rows = self.expect_rows + branch_workload.churn_cursor = self.churn_cursor + return branch_workload + def reconfigure(self) -> None: """ Request the endpoint to reconfigure based on location reported by storage controller diff --git a/test_runner/regress/test_compaction.py b/test_runner/regress/test_compaction.py index fe0422088a..d0a2349ccf 100644 --- a/test_runner/regress/test_compaction.py +++ b/test_runner/regress/test_compaction.py @@ -112,7 +112,11 @@ page_cache_size=10 @skip_in_debug_build("only run with release build") -def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder): +@pytest.mark.parametrize( + "with_branches", + ["with_branches", "no_branches"], +) +def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder, with_branches: str): SMOKE_CONF = { # Run both gc and gc-compaction. "gc_period": "5s", @@ -143,12 +147,17 @@ def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder): log.info("Writing initial data ...") workload.write_rows(row_count, env.pageserver.id) + child_workloads: list[Workload] = [] + for i in range(1, churn_rounds + 1): if i % 10 == 0: log.info(f"Running churn round {i}/{churn_rounds} ...") - - if (i - 1) % 10 == 0: - # Run gc-compaction every 10 rounds to ensure the test doesn't take too long time. + if i % 10 == 5 and with_branches == "with_branches": + branch_name = f"child-{i}" + branch_timeline_id = env.create_branch(branch_name) + child_workloads.append(workload.branch(branch_timeline_id, branch_name)) + if (i - 1) % 10 == 0 or (i - 1) % 10 == 1: + # Run gc-compaction twice every 10 rounds to ensure the test doesn't take too long time. ps_http.timeline_compact( tenant_id, timeline_id, @@ -179,6 +188,9 @@ def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder): log.info("Validating at workload end ...") workload.validate(env.pageserver.id) + for child_workload in child_workloads: + log.info(f"Validating at branch {child_workload.branch_name}") + child_workload.validate(env.pageserver.id) # Run a legacy compaction+gc to ensure gc-compaction can coexist with legacy compaction. ps_http.timeline_checkpoint(tenant_id, timeline_id, wait_until_uploaded=True) From c7429af8a0765b516bd3f6a49c8bc6632deffbf7 Mon Sep 17 00:00:00 2001 From: Mikhail Kot Date: Wed, 15 Jan 2025 22:29:18 +0000 Subject: [PATCH 23/31] Enable dblink (#10358) Update compute image to include dblink #3720 --- compute/compute-node.Dockerfile | 1 + 1 file changed, 1 insertion(+) diff --git a/compute/compute-node.Dockerfile b/compute/compute-node.Dockerfile index 1ee159e5df..f56a8358d2 100644 --- a/compute/compute-node.Dockerfile +++ b/compute/compute-node.Dockerfile @@ -66,6 +66,7 @@ RUN cd postgres && \ make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s -C src/interfaces/libpq install && \ # Enable some of contrib extensions echo 'trusted = true' >> /usr/local/pgsql/share/extension/autoinc.control && \ + echo 'trusted = true' >> /usr/local/pgsql/share/extension/dblink.control && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/bloom.control && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/earthdistance.control && \ echo 'trusted = true' >> /usr/local/pgsql/share/extension/insert_username.control && \ From 2eda484ef6dac15e0e7a72955e9d194ccfe3326f Mon Sep 17 00:00:00 2001 From: Matthias van de Meent Date: Thu, 16 Jan 2025 03:43:47 +0100 Subject: [PATCH 24/31] prefetch: Read more frequently from TCP buffer (#10394) This reduces pressure on the OS TCP read buffer by increasing the moments we read data out of the receive buffer, and increasing the number of bytes we can pull from that buffer when we do reads. ## Problem A backend may not always consume its prefetch data quick enough ## Summary of changes We add a new function `prefetch_pump_state` which pulls as many prefetch requests from the OS TCP receive buffer as possible, but without blocking. This thus reduces pressure on OS-level TCP buffers, thus increasing throughput by limiting throttling caused by full TCP buffers. --- pgxn/neon/libpagestore.c | 70 +++++++++++++++++++++++++++++++++++- pgxn/neon/pagestore_client.h | 20 +++++++++++ pgxn/neon/pagestore_smgr.c | 66 ++++++++++++++++++++++++++++++++++ 3 files changed, 155 insertions(+), 1 deletion(-) diff --git a/pgxn/neon/libpagestore.c b/pgxn/neon/libpagestore.c index 769befb4e5..4460e3b40c 100644 --- a/pgxn/neon/libpagestore.c +++ b/pgxn/neon/libpagestore.c @@ -911,7 +911,74 @@ pageserver_receive(shardno_t shard_no) } PG_CATCH(); { - neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due malformatted response"); + neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due to failure while parsing response"); + pageserver_disconnect(shard_no); + PG_RE_THROW(); + } + PG_END_TRY(); + + if (message_level_is_interesting(PageStoreTrace)) + { + char *msg = nm_to_string((NeonMessage *) resp); + + neon_shard_log(shard_no, PageStoreTrace, "got response: %s", msg); + pfree(msg); + } + } + else if (rc == -1) + { + neon_shard_log(shard_no, LOG, "pageserver_receive disconnect: psql end of copy data: %s", pchomp(PQerrorMessage(pageserver_conn))); + pageserver_disconnect(shard_no); + resp = NULL; + } + else if (rc == -2) + { + char *msg = pchomp(PQerrorMessage(pageserver_conn)); + + pageserver_disconnect(shard_no); + neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: could not read COPY data: %s", msg); + } + else + { + pageserver_disconnect(shard_no); + neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc); + } + + shard->nresponses_received++; + return (NeonResponse *) resp; +} + +static NeonResponse * +pageserver_try_receive(shardno_t shard_no) +{ + StringInfoData resp_buff; + NeonResponse *resp; + PageServer *shard = &page_servers[shard_no]; + PGconn *pageserver_conn = shard->conn; + /* read response */ + int rc; + + if (shard->state != PS_Connected) + return NULL; + + Assert(pageserver_conn); + + rc = PQgetCopyData(shard->conn, &resp_buff.data, 1 /* async = true */); + + if (rc == 0) + return NULL; + else if (rc > 0) + { + PG_TRY(); + { + resp_buff.len = rc; + resp_buff.cursor = 0; + resp = nm_unpack_response(&resp_buff); + PQfreemem(resp_buff.data); + } + PG_CATCH(); + { + neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due to failure while parsing response"); pageserver_disconnect(shard_no); PG_RE_THROW(); } @@ -980,6 +1047,7 @@ page_server_api api = .send = pageserver_send, .flush = pageserver_flush, .receive = pageserver_receive, + .try_receive = pageserver_try_receive, .disconnect = pageserver_disconnect_shard }; diff --git a/pgxn/neon/pagestore_client.h b/pgxn/neon/pagestore_client.h index 37bc4f7886..b751235595 100644 --- a/pgxn/neon/pagestore_client.h +++ b/pgxn/neon/pagestore_client.h @@ -192,9 +192,29 @@ typedef uint16 shardno_t; typedef struct { + /* + * Send this request to the PageServer associated with this shard. + */ bool (*send) (shardno_t shard_no, NeonRequest * request); + /* + * Blocking read for the next response of this shard. + * + * When a CANCEL signal is handled, the connection state will be + * unmodified. + */ NeonResponse *(*receive) (shardno_t shard_no); + /* + * Try get the next response from the TCP buffers, if any. + * Returns NULL when the data is not yet available. + */ + NeonResponse *(*try_receive) (shardno_t shard_no); + /* + * Make sure all requests are sent to PageServer. + */ bool (*flush) (shardno_t shard_no); + /* + * Disconnect from this pageserver shard. + */ void (*disconnect) (shardno_t shard_no); } page_server_api; diff --git a/pgxn/neon/pagestore_smgr.c b/pgxn/neon/pagestore_smgr.c index 7a4c0ef487..54cacea984 100644 --- a/pgxn/neon/pagestore_smgr.c +++ b/pgxn/neon/pagestore_smgr.c @@ -405,6 +405,56 @@ compact_prefetch_buffers(void) return false; } +/* + * If there might be responses still in the TCP buffer, then + * we should try to use those, so as to reduce any TCP backpressure + * on the OS/PS side. + * + * This procedure handles that. + * + * Note that this is only valid as long as the only pipelined + * operations in the TCP buffer are getPage@Lsn requests. + */ +static void +prefetch_pump_state(void) +{ + while (MyPState->ring_receive != MyPState->ring_flush) + { + NeonResponse *response; + PrefetchRequest *slot; + MemoryContext old; + + slot = GetPrfSlot(MyPState->ring_receive); + + old = MemoryContextSwitchTo(MyPState->errctx); + response = page_server->try_receive(slot->shard_no); + MemoryContextSwitchTo(old); + + if (response == NULL) + break; + + /* The slot should still be valid */ + if (slot->status != PRFS_REQUESTED || + slot->response != NULL || + slot->my_ring_index != MyPState->ring_receive) + neon_shard_log(slot->shard_no, ERROR, + "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu", + slot->status, slot->response, + (long) slot->my_ring_index, (long) MyPState->ring_receive); + + /* update prefetch state */ + MyPState->n_responses_buffered += 1; + MyPState->n_requests_inflight -= 1; + MyPState->ring_receive += 1; + MyNeonCounters->getpage_prefetches_buffered = + MyPState->n_responses_buffered; + + /* update slot state */ + slot->status = PRFS_RECEIVED; + slot->response = response; + } +} + void readahead_buffer_resize(int newsize, void *extra) { @@ -2808,6 +2858,8 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, MyPState->ring_last <= ring_index); } + prefetch_pump_state(); + return false; } @@ -2849,6 +2901,8 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum) Assert(ring_index < MyPState->ring_unused && MyPState->ring_last <= ring_index); + prefetch_pump_state(); + return false; } #endif /* PG_MAJORVERSION_NUM < 17 */ @@ -2891,6 +2945,8 @@ neon_writeback(SMgrRelation reln, ForkNumber forknum, */ neon_log(SmgrTrace, "writeback noop"); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (IS_LOCAL_REL(reln)) mdwriteback(reln, forknum, blocknum, nblocks); @@ -3145,6 +3201,8 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1, NULL); neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsns, buffer); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln)) { @@ -3282,6 +3340,8 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, neon_read_at_lsnv(InfoFromSMgrRel(reln), forknum, blocknum, request_lsns, buffers, nblocks, read); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln)) { @@ -3450,6 +3510,8 @@ neon_write(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, const vo lfc_write(InfoFromSMgrRel(reln), forknum, blocknum, buffer); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (IS_LOCAL_REL(reln)) #if PG_MAJORVERSION_NUM >= 17 @@ -3503,6 +3565,8 @@ neon_writev(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno, lfc_writev(InfoFromSMgrRel(reln), forknum, blkno, buffers, nblocks); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (IS_LOCAL_REL(reln)) mdwritev(reln, forknum, blocknum, &buffer, 1, skipFsync); @@ -3792,6 +3856,8 @@ neon_immedsync(SMgrRelation reln, ForkNumber forknum) neon_log(SmgrTrace, "[NEON_SMGR] immedsync noop"); + prefetch_pump_state(); + #ifdef DEBUG_COMPARE_LOCAL if (IS_LOCAL_REL(reln)) mdimmedsync(reln, forknum); From 6fe4c6798f76aaf103f1f44166970f4677ff6c9d Mon Sep 17 00:00:00 2001 From: Arseny Sher Date: Thu, 16 Jan 2025 11:01:19 +0300 Subject: [PATCH 25/31] Add START_WAL_PUSH proto_version and allow_timeline_creation options. (#10406) ## Problem As part of https://github.com/neondatabase/neon/issues/8614 we need to pass options to START_WAL_PUSH. ## Summary of changes Add two options. `allow_timeline_creation`, default true, disables implicit timeline creation in the connection from compute. Eventually such creation will be forbidden completely, but as we migrate to configurations we need to support both: current mode and configurations enabled where creation by compute is disabled. `proto_version` specifies compute <-> sk protocol version. We have it currently in the first greeting package also, but I plan to change tag size from u64 to u8, which would make it hard to use. Command is more appropriate place for it anyway. --- safekeeper/src/handler.rs | 107 ++++++++++++++++-- safekeeper/src/receive_wal.rs | 57 +++++++--- safekeeper/src/safekeeper.rs | 11 +- .../tests/walproposer_sim/safekeeper.rs | 6 +- 4 files changed, 152 insertions(+), 29 deletions(-) diff --git a/safekeeper/src/handler.rs b/safekeeper/src/handler.rs index bb639bfb32..e77eeb4130 100644 --- a/safekeeper/src/handler.rs +++ b/safekeeper/src/handler.rs @@ -52,16 +52,70 @@ pub struct SafekeeperPostgresHandler { /// Parsed Postgres command. enum SafekeeperPostgresCommand { - StartWalPush, - StartReplication { start_lsn: Lsn, term: Option }, + StartWalPush { + proto_version: u32, + // Eventually timelines will be always created explicitly by storcon. + // This option allows legacy behaviour for compute to do that until we + // fully migrate. + allow_timeline_creation: bool, + }, + StartReplication { + start_lsn: Lsn, + term: Option, + }, IdentifySystem, TimelineStatus, - JSONCtrl { cmd: AppendLogicalMessage }, + JSONCtrl { + cmd: AppendLogicalMessage, + }, } fn parse_cmd(cmd: &str) -> anyhow::Result { if cmd.starts_with("START_WAL_PUSH") { - Ok(SafekeeperPostgresCommand::StartWalPush) + // Allow additional options in postgres START_REPLICATION style like + // START_WAL_PUSH (proto_version '3', allow_timeline_creation 'false'). + // Parsing here is very naive and breaks in case of commas or + // whitespaces in values, but enough for our purposes. + let re = Regex::new(r"START_WAL_PUSH(\s+?\((.*)\))?").unwrap(); + let caps = re + .captures(cmd) + .context(format!("failed to parse START_WAL_PUSH command {}", cmd))?; + // capture () content + let options = caps.get(2).map(|m| m.as_str()).unwrap_or(""); + // default values + let mut proto_version = 2; + let mut allow_timeline_creation = true; + for kvstr in options.split(",") { + if kvstr.is_empty() { + continue; + } + let mut kvit = kvstr.split_whitespace(); + let key = kvit.next().context(format!( + "failed to parse key in kv {} in command {}", + kvstr, cmd + ))?; + let value = kvit.next().context(format!( + "failed to parse value in kv {} in command {}", + kvstr, cmd + ))?; + let value_trimmed = value.trim_matches('\''); + if key == "proto_version" { + proto_version = value_trimmed.parse::().context(format!( + "failed to parse proto_version value {} in command {}", + value, cmd + ))?; + } + if key == "allow_timeline_creation" { + allow_timeline_creation = value_trimmed.parse::().context(format!( + "failed to parse allow_timeline_creation value {} in command {}", + value, cmd + ))?; + } + } + Ok(SafekeeperPostgresCommand::StartWalPush { + proto_version, + allow_timeline_creation, + }) } else if cmd.starts_with("START_REPLICATION") { let re = Regex::new( // We follow postgres START_REPLICATION LOGICAL options to pass term. @@ -95,7 +149,7 @@ fn parse_cmd(cmd: &str) -> anyhow::Result { fn cmd_to_string(cmd: &SafekeeperPostgresCommand) -> &str { match cmd { - SafekeeperPostgresCommand::StartWalPush => "START_WAL_PUSH", + SafekeeperPostgresCommand::StartWalPush { .. } => "START_WAL_PUSH", SafekeeperPostgresCommand::StartReplication { .. } => "START_REPLICATION", SafekeeperPostgresCommand::TimelineStatus => "TIMELINE_STATUS", SafekeeperPostgresCommand::IdentifySystem => "IDENTIFY_SYSTEM", @@ -293,8 +347,11 @@ impl postgres_backend::Handler self.ttid = TenantTimelineId::new(tenant_id, timeline_id); match cmd { - SafekeeperPostgresCommand::StartWalPush => { - self.handle_start_wal_push(pgb) + SafekeeperPostgresCommand::StartWalPush { + proto_version, + allow_timeline_creation, + } => { + self.handle_start_wal_push(pgb, proto_version, allow_timeline_creation) .instrument(info_span!("WAL receiver")) .await } @@ -467,3 +524,39 @@ impl SafekeeperPostgresHandler { } } } + +#[cfg(test)] +mod tests { + use super::SafekeeperPostgresCommand; + + /// Test parsing of START_WAL_PUSH command + #[test] + fn test_start_wal_push_parse() { + let cmd = "START_WAL_PUSH"; + let parsed = super::parse_cmd(cmd).expect("failed to parse"); + match parsed { + SafekeeperPostgresCommand::StartWalPush { + proto_version, + allow_timeline_creation, + } => { + assert_eq!(proto_version, 2); + assert!(allow_timeline_creation); + } + _ => panic!("unexpected command"), + } + + let cmd = + "START_WAL_PUSH (proto_version '3', allow_timeline_creation 'false', unknown 'hoho')"; + let parsed = super::parse_cmd(cmd).expect("failed to parse"); + match parsed { + SafekeeperPostgresCommand::StartWalPush { + proto_version, + allow_timeline_creation, + } => { + assert_eq!(proto_version, 3); + assert!(!allow_timeline_creation); + } + _ => panic!("unexpected command"), + } + } +} diff --git a/safekeeper/src/receive_wal.rs b/safekeeper/src/receive_wal.rs index daaa8a253d..cb42f6f414 100644 --- a/safekeeper/src/receive_wal.rs +++ b/safekeeper/src/receive_wal.rs @@ -200,9 +200,14 @@ impl SafekeeperPostgresHandler { pub async fn handle_start_wal_push( &mut self, pgb: &mut PostgresBackend, + proto_version: u32, + allow_timeline_creation: bool, ) -> Result<(), QueryError> { let mut tli: Option = None; - if let Err(end) = self.handle_start_wal_push_guts(pgb, &mut tli).await { + if let Err(end) = self + .handle_start_wal_push_guts(pgb, &mut tli, proto_version, allow_timeline_creation) + .await + { // Log the result and probably send it to the client, closing the stream. let handle_end_fut = pgb.handle_copy_stream_end(end); // If we managed to create the timeline, augment logging with current LSNs etc. @@ -222,6 +227,8 @@ impl SafekeeperPostgresHandler { &mut self, pgb: &mut PostgresBackend, tli: &mut Option, + proto_version: u32, + allow_timeline_creation: bool, ) -> Result<(), CopyStreamHandlerEnd> { // The `tli` parameter is only used for passing _out_ a timeline, one should // not have been passed in. @@ -250,12 +257,17 @@ impl SafekeeperPostgresHandler { conn_id: self.conn_id, pgb_reader: &mut pgb_reader, peer_addr, + proto_version, acceptor_handle: &mut acceptor_handle, global_timelines: self.global_timelines.clone(), }; - // Read first message and create timeline if needed. - let res = network_reader.read_first_message().await; + // Read first message and create timeline if needed and allowed. This + // won't be when timelines will be always created by storcon and + // allow_timeline_creation becomes false. + let res = network_reader + .read_first_message(allow_timeline_creation) + .await; let network_res = if let Ok((timeline, next_msg)) = res { let pageserver_feedback_rx: tokio::sync::broadcast::Receiver = @@ -313,6 +325,7 @@ struct NetworkReader<'a, IO> { conn_id: ConnectionId, pgb_reader: &'a mut PostgresBackendReader, peer_addr: SocketAddr, + proto_version: u32, // WalAcceptor is spawned when we learn server info from walproposer and // create timeline; handle is put here. acceptor_handle: &'a mut Option>>, @@ -322,9 +335,10 @@ struct NetworkReader<'a, IO> { impl NetworkReader<'_, IO> { async fn read_first_message( &mut self, + allow_timeline_creation: bool, ) -> Result<(WalResidentTimeline, ProposerAcceptorMessage), CopyStreamHandlerEnd> { // Receive information about server to create timeline, if not yet. - let next_msg = read_message(self.pgb_reader).await?; + let next_msg = read_message(self.pgb_reader, self.proto_version).await?; let tli = match next_msg { ProposerAcceptorMessage::Greeting(ref greeting) => { info!( @@ -336,17 +350,22 @@ impl NetworkReader<'_, IO> { system_id: greeting.system_id, wal_seg_size: greeting.wal_seg_size, }; - let tli = self - .global_timelines - .create( - self.ttid, - Configuration::empty(), - server_info, - Lsn::INVALID, - Lsn::INVALID, - ) - .await - .context("create timeline")?; + let tli = if allow_timeline_creation { + self.global_timelines + .create( + self.ttid, + Configuration::empty(), + server_info, + Lsn::INVALID, + Lsn::INVALID, + ) + .await + .context("create timeline")? + } else { + self.global_timelines + .get(self.ttid) + .context("get timeline")? + }; tli.wal_residence_guard().await? } _ => { @@ -375,7 +394,7 @@ impl NetworkReader<'_, IO> { )); // Forward all messages to WalAcceptor - read_network_loop(self.pgb_reader, msg_tx, next_msg).await + read_network_loop(self.pgb_reader, msg_tx, next_msg, self.proto_version).await } } @@ -383,9 +402,10 @@ impl NetworkReader<'_, IO> { /// TODO: Return Ok(None) on graceful termination. async fn read_message( pgb_reader: &mut PostgresBackendReader, + proto_version: u32, ) -> Result { let copy_data = pgb_reader.read_copy_message().await?; - let msg = ProposerAcceptorMessage::parse(copy_data)?; + let msg = ProposerAcceptorMessage::parse(copy_data, proto_version)?; Ok(msg) } @@ -393,6 +413,7 @@ async fn read_network_loop( pgb_reader: &mut PostgresBackendReader, msg_tx: Sender, mut next_msg: ProposerAcceptorMessage, + proto_version: u32, ) -> Result<(), CopyStreamHandlerEnd> { /// Threshold for logging slow WalAcceptor sends. const SLOW_THRESHOLD: Duration = Duration::from_secs(5); @@ -425,7 +446,7 @@ async fn read_network_loop( WAL_RECEIVER_QUEUE_DEPTH_TOTAL.inc(); WAL_RECEIVER_QUEUE_SIZE_TOTAL.add(size as i64); - next_msg = read_message(pgb_reader).await?; + next_msg = read_message(pgb_reader, proto_version).await?; } } diff --git a/safekeeper/src/safekeeper.rs b/safekeeper/src/safekeeper.rs index 06403228e9..45e19c31b6 100644 --- a/safekeeper/src/safekeeper.rs +++ b/safekeeper/src/safekeeper.rs @@ -29,7 +29,7 @@ use utils::{ lsn::Lsn, }; -const SK_PROTOCOL_VERSION: u32 = 2; +pub const SK_PROTOCOL_VERSION: u32 = 2; pub const UNKNOWN_SERVER_VERSION: u32 = 0; #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, PartialOrd, Ord)] @@ -317,7 +317,14 @@ pub enum ProposerAcceptorMessage { impl ProposerAcceptorMessage { /// Parse proposer message. - pub fn parse(msg_bytes: Bytes) -> Result { + pub fn parse(msg_bytes: Bytes, proto_version: u32) -> Result { + if proto_version != SK_PROTOCOL_VERSION { + bail!( + "incompatible protocol version {}, expected {}", + proto_version, + SK_PROTOCOL_VERSION + ); + } // xxx using Reader is inefficient but easy to work with bincode let mut stream = msg_bytes.reader(); // u64 is here to avoid padding; it will be removed once we stop packing C structs into the wire as is diff --git a/safekeeper/tests/walproposer_sim/safekeeper.rs b/safekeeper/tests/walproposer_sim/safekeeper.rs index e0d593851e..0023a4d22a 100644 --- a/safekeeper/tests/walproposer_sim/safekeeper.rs +++ b/safekeeper/tests/walproposer_sim/safekeeper.rs @@ -15,7 +15,9 @@ use desim::{ }; use http::Uri; use safekeeper::{ - safekeeper::{ProposerAcceptorMessage, SafeKeeper, UNKNOWN_SERVER_VERSION}, + safekeeper::{ + ProposerAcceptorMessage, SafeKeeper, SK_PROTOCOL_VERSION, UNKNOWN_SERVER_VERSION, + }, state::{TimelinePersistentState, TimelineState}, timeline::TimelineError, wal_storage::Storage, @@ -285,7 +287,7 @@ impl ConnState { bail!("finished processing START_REPLICATION") } - let msg = ProposerAcceptorMessage::parse(copy_data)?; + let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTOCOL_VERSION)?; debug!("got msg: {:?}", msg); self.process(msg, global) } else { From 7be971081aee245b0b78ff22766d8eeb020b66f2 Mon Sep 17 00:00:00 2001 From: Matthias van de Meent Date: Thu, 16 Jan 2025 09:34:11 +0100 Subject: [PATCH 26/31] Make sure we request pages with a known-flushed LSN. (#10413) This should fix the largest source of flakyness of test_nbtree_pagesplit_cycleid. ## Problem https://github.com/neondatabase/neon/issues/10390 ## Summary of changes By using a guaranteed-flushed LSN, we ensure that PS won't have to wait forever. (If it does wait forever, we know the issue can't be with Compute's WAL) --- .../regress/test_nbtree_pagesplit_cycleid.py | 20 ++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-) diff --git a/test_runner/regress/test_nbtree_pagesplit_cycleid.py b/test_runner/regress/test_nbtree_pagesplit_cycleid.py index 558557aeba..32ec6fcb92 100644 --- a/test_runner/regress/test_nbtree_pagesplit_cycleid.py +++ b/test_runner/regress/test_nbtree_pagesplit_cycleid.py @@ -4,9 +4,19 @@ import time from fixtures.neon_fixtures import NeonEnv BTREE_NUM_CYCLEID_PAGES = """ - WITH raw_pages AS ( - SELECT blkno, get_raw_page_at_lsn('t_uidx', 'main', blkno, NULL, NULL) page - FROM generate_series(1, pg_relation_size('t_uidx'::regclass) / 8192) blkno + WITH lsns AS ( + /* + * pg_switch_wal() ensures we have an LSN that + * 1. is after any previous modifications, but also, + * 2. (critically) is flushed, preventing any issues with waiting for + * unflushed WAL in PageServer. + */ + SELECT pg_switch_wal() as lsn + ), + raw_pages AS ( + SELECT blkno, get_raw_page_at_lsn('t_uidx', 'main', blkno, lsn, lsn) page + FROM generate_series(1, pg_relation_size('t_uidx'::regclass) / 8192) AS blkno, + lsns l(lsn) ), parsed_pages AS ( /* cycle ID is the last 2 bytes of the btree page */ @@ -36,7 +46,6 @@ def test_nbtree_pagesplit_cycleid(neon_simple_env: NeonEnv): ses1.execute("CREATE UNIQUE INDEX t_uidx ON t(id);") ses1.execute("INSERT INTO t (txt) SELECT i::text FROM generate_series(1, 2035) i;") - ses1.execute("SELECT neon_xlogflush();") ses1.execute(BTREE_NUM_CYCLEID_PAGES) pages = ses1.fetchall() assert ( @@ -57,7 +66,6 @@ def test_nbtree_pagesplit_cycleid(neon_simple_env: NeonEnv): ses1.execute("DELETE FROM t WHERE id <= 610;") # Flush wal, for checking purposes - ses1.execute("SELECT neon_xlogflush();") ses1.execute(BTREE_NUM_CYCLEID_PAGES) pages = ses1.fetchall() assert len(pages) == 0, f"No back splits with cycle ID expected, got batches of {pages} instead" @@ -108,8 +116,6 @@ def test_nbtree_pagesplit_cycleid(neon_simple_env: NeonEnv): # unpin the btree page, allowing s3's vacuum to complete ses2.execute("FETCH ALL FROM foo;") ses2.execute("ROLLBACK;") - # flush WAL to make sure PS is up-to-date - ses1.execute("SELECT neon_xlogflush();") # check that our expectations are correct ses1.execute(BTREE_NUM_CYCLEID_PAGES) pages = ses1.fetchall() From 58f6af6c9ad5bbc6238999769c801cdc52495bac Mon Sep 17 00:00:00 2001 From: Tristan Partin Date: Thu, 16 Jan 2025 02:35:36 -0600 Subject: [PATCH 27/31] Clean up compute_ctl extension server code (#10417) --- compute_tools/src/http/routes/extension_server.rs | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/compute_tools/src/http/routes/extension_server.rs b/compute_tools/src/http/routes/extension_server.rs index ee5bc675ba..5cc9b6d277 100644 --- a/compute_tools/src/http/routes/extension_server.rs +++ b/compute_tools/src/http/routes/extension_server.rs @@ -17,7 +17,8 @@ use crate::{ #[derive(Debug, Clone, Deserialize)] pub(in crate::http) struct ExtensionServerParams { - is_library: Option, + #[serde(default)] + is_library: bool, } /// Download a remote extension. @@ -51,7 +52,7 @@ pub(in crate::http) async fn download_extension( remote_extensions.get_ext( &filename, - params.is_library.unwrap_or(false), + params.is_library, &compute.build_tag, &compute.pgversion, ) From 86dbc44db12d1ebb5b9532a270fcfc68caffd8d1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?JC=20Gr=C3=BCnhage?= Date: Thu, 16 Jan 2025 10:20:24 +0100 Subject: [PATCH 28/31] CI: Run check-codestyle-rust as part of pre-merge-checks (#10387) ## Problem When multiple changes are grouped in a merge group to be merged as part of the merge queue, the changes might individually pass `check-codestyle-rust` but not in their combined form. ## Summary of changes - Move `check-codestyle-rust` into a reusable workflow that is called from it's previous location in `build_and_test.yml`, and additionally call it from `pre_merge_checks.yml`. The additional call does not run on ARM, only x86, to ensure the merge queue continues being responsive. - Trigger `pre_merge_checks.yml` on PRs that change any of the workflows running in `pre_merge_checks.yml`, so that we get feedback on those early an not only after trying to merge those changes. --- .github/workflows/_check-codestyle-rust.yml | 91 +++++++++++++++++++++ .github/workflows/build_and_test.yml | 76 ++--------------- .github/workflows/pre-merge-checks.yml | 33 ++++++++ 3 files changed, 129 insertions(+), 71 deletions(-) create mode 100644 .github/workflows/_check-codestyle-rust.yml diff --git a/.github/workflows/_check-codestyle-rust.yml b/.github/workflows/_check-codestyle-rust.yml new file mode 100644 index 0000000000..cbc47c6406 --- /dev/null +++ b/.github/workflows/_check-codestyle-rust.yml @@ -0,0 +1,91 @@ +name: Check Codestyle Rust + +on: + workflow_call: + inputs: + build-tools-image: + description: "build-tools image" + required: true + type: string + archs: + description: "Json array of architectures to run on" + type: string + + +defaults: + run: + shell: bash -euxo pipefail {0} + +jobs: + check-codestyle-rust: + strategy: + matrix: + arch: ${{ fromJson(inputs.archs) }} + runs-on: ${{ fromJson(format('["self-hosted", "{0}"]', matrix.arch == 'arm64' && 'small-arm64' || 'small')) }} + + container: + image: ${{ inputs.build-tools-image }} + credentials: + username: ${{ secrets.NEON_DOCKERHUB_USERNAME }} + password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }} + options: --init + + steps: + - name: Checkout + uses: actions/checkout@v4 + with: + submodules: true + + - name: Cache cargo deps + uses: actions/cache@v4 + with: + path: | + ~/.cargo/registry + !~/.cargo/registry/src + ~/.cargo/git + target + key: v1-${{ runner.os }}-${{ runner.arch }}-cargo-${{ hashFiles('./Cargo.lock') }}-${{ hashFiles('./rust-toolchain.toml') }}-rust + + # Some of our rust modules use FFI and need those to be checked + - name: Get postgres headers + run: make postgres-headers -j$(nproc) + + # cargo hack runs the given cargo subcommand (clippy in this case) for all feature combinations. + # This will catch compiler & clippy warnings in all feature combinations. + # TODO: use cargo hack for build and test as well, but, that's quite expensive. + # NB: keep clippy args in sync with ./run_clippy.sh + # + # The only difference between "clippy --debug" and "clippy --release" is that in --release mode, + # #[cfg(debug_assertions)] blocks are not built. It's not worth building everything for second + # time just for that, so skip "clippy --release". + - run: | + CLIPPY_COMMON_ARGS="$( source .neon_clippy_args; echo "$CLIPPY_COMMON_ARGS")" + if [ "$CLIPPY_COMMON_ARGS" = "" ]; then + echo "No clippy args found in .neon_clippy_args" + exit 1 + fi + echo "CLIPPY_COMMON_ARGS=${CLIPPY_COMMON_ARGS}" >> $GITHUB_ENV + - name: Run cargo clippy (debug) + run: cargo hack --features default --ignore-unknown-features --feature-powerset clippy $CLIPPY_COMMON_ARGS + + - name: Check documentation generation + run: cargo doc --workspace --no-deps --document-private-items + env: + RUSTDOCFLAGS: "-Dwarnings -Arustdoc::private_intra_doc_links" + + # Use `${{ !cancelled() }}` to run quck tests after the longer clippy run + - name: Check formatting + if: ${{ !cancelled() }} + run: cargo fmt --all -- --check + + # https://github.com/facebookincubator/cargo-guppy/tree/bec4e0eb29dcd1faac70b1b5360267fc02bf830e/tools/cargo-hakari#2-keep-the-workspace-hack-up-to-date-in-ci + - name: Check rust dependencies + if: ${{ !cancelled() }} + run: | + cargo hakari generate --diff # workspace-hack Cargo.toml is up-to-date + cargo hakari manage-deps --dry-run # all workspace crates depend on workspace-hack + + # https://github.com/EmbarkStudios/cargo-deny + - name: Check rust licenses/bans/advisories/sources + if: ${{ !cancelled() }} + run: cargo deny check --hide-inclusion-graph diff --git a/.github/workflows/build_and_test.yml b/.github/workflows/build_and_test.yml index 489a93f46d..9ec5273af7 100644 --- a/.github/workflows/build_and_test.yml +++ b/.github/workflows/build_and_test.yml @@ -164,77 +164,11 @@ jobs: check-codestyle-rust: needs: [ check-permissions, build-build-tools-image ] - strategy: - matrix: - arch: [ x64, arm64 ] - runs-on: ${{ fromJson(format('["self-hosted", "{0}"]', matrix.arch == 'arm64' && 'small-arm64' || 'small')) }} - - container: - image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm - credentials: - username: ${{ secrets.NEON_DOCKERHUB_USERNAME }} - password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }} - options: --init - - steps: - - name: Checkout - uses: actions/checkout@v4 - with: - submodules: true - - - name: Cache cargo deps - uses: actions/cache@v4 - with: - path: | - ~/.cargo/registry - !~/.cargo/registry/src - ~/.cargo/git - target - key: v1-${{ runner.os }}-${{ runner.arch }}-cargo-${{ hashFiles('./Cargo.lock') }}-${{ hashFiles('./rust-toolchain.toml') }}-rust - - # Some of our rust modules use FFI and need those to be checked - - name: Get postgres headers - run: make postgres-headers -j$(nproc) - - # cargo hack runs the given cargo subcommand (clippy in this case) for all feature combinations. - # This will catch compiler & clippy warnings in all feature combinations. - # TODO: use cargo hack for build and test as well, but, that's quite expensive. - # NB: keep clippy args in sync with ./run_clippy.sh - # - # The only difference between "clippy --debug" and "clippy --release" is that in --release mode, - # #[cfg(debug_assertions)] blocks are not built. It's not worth building everything for second - # time just for that, so skip "clippy --release". - - run: | - CLIPPY_COMMON_ARGS="$( source .neon_clippy_args; echo "$CLIPPY_COMMON_ARGS")" - if [ "$CLIPPY_COMMON_ARGS" = "" ]; then - echo "No clippy args found in .neon_clippy_args" - exit 1 - fi - echo "CLIPPY_COMMON_ARGS=${CLIPPY_COMMON_ARGS}" >> $GITHUB_ENV - - name: Run cargo clippy (debug) - run: cargo hack --features default --ignore-unknown-features --feature-powerset clippy $CLIPPY_COMMON_ARGS - - - name: Check documentation generation - run: cargo doc --workspace --no-deps --document-private-items - env: - RUSTDOCFLAGS: "-Dwarnings -Arustdoc::private_intra_doc_links" - - # Use `${{ !cancelled() }}` to run quck tests after the longer clippy run - - name: Check formatting - if: ${{ !cancelled() }} - run: cargo fmt --all -- --check - - # https://github.com/facebookincubator/cargo-guppy/tree/bec4e0eb29dcd1faac70b1b5360267fc02bf830e/tools/cargo-hakari#2-keep-the-workspace-hack-up-to-date-in-ci - - name: Check rust dependencies - if: ${{ !cancelled() }} - run: | - cargo hakari generate --diff # workspace-hack Cargo.toml is up-to-date - cargo hakari manage-deps --dry-run # all workspace crates depend on workspace-hack - - # https://github.com/EmbarkStudios/cargo-deny - - name: Check rust licenses/bans/advisories/sources - if: ${{ !cancelled() }} - run: cargo deny check --hide-inclusion-graph + uses: ./.github/workflows/_check-codestyle-rust.yml + with: + build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm + archs: '["x64", "arm64"]' + secrets: inherit build-and-test-locally: needs: [ tag, build-build-tools-image ] diff --git a/.github/workflows/pre-merge-checks.yml b/.github/workflows/pre-merge-checks.yml index b2e00d94f7..e6dfbaeed8 100644 --- a/.github/workflows/pre-merge-checks.yml +++ b/.github/workflows/pre-merge-checks.yml @@ -1,6 +1,12 @@ name: Pre-merge checks on: + pull_request: + paths: + - .github/workflows/_check-codestyle-python.yml + - .github/workflows/_check-codestyle-rust.yml + - .github/workflows/build-build-tools-image.yml + - .github/workflows/pre-merge-checks.yml merge_group: branches: - main @@ -17,8 +23,10 @@ jobs: runs-on: ubuntu-22.04 outputs: python-changed: ${{ steps.python-src.outputs.any_changed }} + rust-changed: ${{ steps.rust-src.outputs.any_changed }} steps: - uses: actions/checkout@v4 + - uses: tj-actions/changed-files@4edd678ac3f81e2dc578756871e4d00c19191daf # v45.0.4 id: python-src with: @@ -30,11 +38,25 @@ jobs: poetry.lock pyproject.toml + - uses: tj-actions/changed-files@4edd678ac3f81e2dc578756871e4d00c19191daf # v45.0.4 + id: rust-src + with: + files: | + .github/workflows/_check-codestyle-rust.yml + .github/workflows/build-build-tools-image.yml + .github/workflows/pre-merge-checks.yml + **/**.rs + **/Cargo.toml + Cargo.toml + Cargo.lock + - name: PRINT ALL CHANGED FILES FOR DEBUG PURPOSES env: PYTHON_CHANGED_FILES: ${{ steps.python-src.outputs.all_changed_files }} + RUST_CHANGED_FILES: ${{ steps.rust-src.outputs.all_changed_files }} run: | echo "${PYTHON_CHANGED_FILES}" + echo "${RUST_CHANGED_FILES}" build-build-tools-image: if: needs.get-changed-files.outputs.python-changed == 'true' @@ -55,6 +77,16 @@ jobs: build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm-x64 secrets: inherit + check-codestyle-rust: + if: needs.get-changed-files.outputs.rust-changed == 'true' + needs: [ get-changed-files, build-build-tools-image ] + uses: ./.github/workflows/_check-codestyle-rust.yml + with: + # `-bookworm-x64` suffix should match the combination in `build-build-tools-image` + build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm-x64 + archs: '["x64"]' + secrets: inherit + # To get items from the merge queue merged into main we need to satisfy "Status checks that are required". # Currently we require 2 jobs (checks with exact name): # - conclusion @@ -67,6 +99,7 @@ jobs: needs: - get-changed-files - check-codestyle-python + - check-codestyle-rust runs-on: ubuntu-22.04 steps: - name: Create fake `neon-cloud-e2e` check From 21d7b6a258e16542cf025790072850e80e40fdff Mon Sep 17 00:00:00 2001 From: John Spray Date: Thu, 16 Jan 2025 14:11:33 +0000 Subject: [PATCH 29/31] tests: refactor test_tenant_delete_races_timeline_creation (#10425) ## Problem Threads spawned in `test_tenant_delete_races_timeline_creation` are not joined before the test ends, and can generate `PytestUnhandledThreadExceptionWarning` in other tests. https://neon-github-public-dev.s3.amazonaws.com/reports/pr-10419/12805365523/index.html#/testresult/53a72568acd04dbd ## Summary of changes - Wrap threads in ThreadPoolExecutor which will join them before the test ends - Remove a spurious deletion call -- the background thread doing deletion ought to succeed. --- test_runner/regress/test_tenant_delete.py | 71 ++++++++++++----------- 1 file changed, 38 insertions(+), 33 deletions(-) diff --git a/test_runner/regress/test_tenant_delete.py b/test_runner/regress/test_tenant_delete.py index 48e55c1ab1..3720f653c5 100644 --- a/test_runner/regress/test_tenant_delete.py +++ b/test_runner/regress/test_tenant_delete.py @@ -1,6 +1,7 @@ from __future__ import annotations import json +from concurrent.futures import ThreadPoolExecutor from threading import Thread import pytest @@ -253,29 +254,8 @@ def test_tenant_delete_races_timeline_creation(neon_env_builder: NeonEnvBuilder) ps_http.configure_failpoints((BEFORE_INITDB_UPLOAD_FAILPOINT, "pause")) def timeline_create(): - try: - ps_http.timeline_create(env.pg_version, tenant_id, TimelineId.generate(), timeout=1) - raise RuntimeError("creation succeeded even though it shouldn't") - except ReadTimeout: - pass - - Thread(target=timeline_create).start() - - def hit_initdb_upload_failpoint(): - env.pageserver.assert_log_contains(f"at failpoint {BEFORE_INITDB_UPLOAD_FAILPOINT}") - - wait_until(hit_initdb_upload_failpoint) - - def creation_connection_timed_out(): - env.pageserver.assert_log_contains( - "POST.*/timeline.* request was dropped before completing" - ) - - # Wait so that we hit the timeout and the connection is dropped - # (But timeline creation still continues) - wait_until(creation_connection_timed_out) - - ps_http.configure_failpoints((DELETE_BEFORE_CLEANUP_FAILPOINT, "pause")) + ps_http.timeline_create(env.pg_version, tenant_id, TimelineId.generate(), timeout=1) + raise RuntimeError("creation succeeded even though it shouldn't") def tenant_delete(): def tenant_delete_inner(): @@ -283,21 +263,46 @@ def test_tenant_delete_races_timeline_creation(neon_env_builder: NeonEnvBuilder) wait_until(tenant_delete_inner) - Thread(target=tenant_delete).start() + # We will spawn background threads for timeline creation and tenant deletion. They will both + # get blocked on our failpoint. + with ThreadPoolExecutor(max_workers=1) as executor: + create_fut = executor.submit(timeline_create) - def deletion_arrived(): - env.pageserver.assert_log_contains( - f"cfg failpoint: {DELETE_BEFORE_CLEANUP_FAILPOINT} pause" - ) + def hit_initdb_upload_failpoint(): + env.pageserver.assert_log_contains(f"at failpoint {BEFORE_INITDB_UPLOAD_FAILPOINT}") - wait_until(deletion_arrived) + wait_until(hit_initdb_upload_failpoint) - ps_http.configure_failpoints((DELETE_BEFORE_CLEANUP_FAILPOINT, "off")) + def creation_connection_timed_out(): + env.pageserver.assert_log_contains( + "POST.*/timeline.* request was dropped before completing" + ) - # Disable the failpoint and wait for deletion to finish - ps_http.configure_failpoints((BEFORE_INITDB_UPLOAD_FAILPOINT, "off")) + # Wait so that we hit the timeout and the connection is dropped + # (But timeline creation still continues) + wait_until(creation_connection_timed_out) - ps_http.tenant_delete(tenant_id) + with pytest.raises(ReadTimeout): + # Our creation failed from the client's point of view. + create_fut.result() + + ps_http.configure_failpoints((DELETE_BEFORE_CLEANUP_FAILPOINT, "pause")) + + delete_fut = executor.submit(tenant_delete) + + def deletion_arrived(): + env.pageserver.assert_log_contains( + f"cfg failpoint: {DELETE_BEFORE_CLEANUP_FAILPOINT} pause" + ) + + wait_until(deletion_arrived) + + ps_http.configure_failpoints((DELETE_BEFORE_CLEANUP_FAILPOINT, "off")) + + # Disable the failpoint and wait for deletion to finish + ps_http.configure_failpoints((BEFORE_INITDB_UPLOAD_FAILPOINT, "off")) + + delete_fut.result() # Physical deletion should have happened assert_prefix_empty( From e436dcad57d22f6d5fcb84ff2118b5f2ec96656f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Arpad=20M=C3=BCller?= Date: Thu, 16 Jan 2025 15:30:49 +0100 Subject: [PATCH 30/31] Rename "disabled" safekeeper scheduling policy to "pause" (#10410) Rename the safekeeper scheduling policy "disabled" to "pause". A rename was requested in https://github.com/neondatabase/neon/pull/10400#discussion_r1916259124, as the "disabled" policy is meant to be analogous to the "pause" policy for pageservers. Also simplify the `SkSchedulingPolicyArg::from_str` function, relying on the `from_str` implementation of `SkSchedulingPolicy`. Latter is used for the database format as well, so it is quite stable. If we ever want to change the UI, we'll need to duplicate the function again but this is cheap. --- control_plane/storcon_cli/src/main.rs | 9 +-------- libs/pageserver_api/src/controller_api.rs | 12 ++++++++---- .../down.sql | 2 ++ .../up.sql | 2 ++ test_runner/regress/test_storage_controller.py | 2 +- 5 files changed, 14 insertions(+), 13 deletions(-) create mode 100644 storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/down.sql create mode 100644 storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/up.sql diff --git a/control_plane/storcon_cli/src/main.rs b/control_plane/storcon_cli/src/main.rs index 96bfad4c86..d9b76b9600 100644 --- a/control_plane/storcon_cli/src/main.rs +++ b/control_plane/storcon_cli/src/main.rs @@ -298,14 +298,7 @@ impl FromStr for SkSchedulingPolicyArg { type Err = anyhow::Error; fn from_str(s: &str) -> Result { - match s { - "active" => Ok(Self(SkSchedulingPolicy::Active)), - "disabled" => Ok(Self(SkSchedulingPolicy::Disabled)), - "decomissioned" => Ok(Self(SkSchedulingPolicy::Decomissioned)), - _ => Err(anyhow::anyhow!( - "Unknown scheduling policy '{s}', try active,disabled,decomissioned" - )), - } + SkSchedulingPolicy::from_str(s).map(Self) } } diff --git a/libs/pageserver_api/src/controller_api.rs b/libs/pageserver_api/src/controller_api.rs index 08d1fa55b9..78e080981a 100644 --- a/libs/pageserver_api/src/controller_api.rs +++ b/libs/pageserver_api/src/controller_api.rs @@ -324,7 +324,7 @@ impl From for String { #[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)] pub enum SkSchedulingPolicy { Active, - Disabled, + Pause, Decomissioned, } @@ -334,9 +334,13 @@ impl FromStr for SkSchedulingPolicy { fn from_str(s: &str) -> Result { Ok(match s { "active" => Self::Active, - "disabled" => Self::Disabled, + "pause" => Self::Pause, "decomissioned" => Self::Decomissioned, - _ => return Err(anyhow::anyhow!("Unknown scheduling state '{s}'")), + _ => { + return Err(anyhow::anyhow!( + "Unknown scheduling policy '{s}', try active,pause,decomissioned" + )) + } }) } } @@ -346,7 +350,7 @@ impl From for String { use SkSchedulingPolicy::*; match value { Active => "active", - Disabled => "disabled", + Pause => "pause", Decomissioned => "decomissioned", } .to_string() diff --git a/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/down.sql b/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/down.sql new file mode 100644 index 0000000000..3c7126e343 --- /dev/null +++ b/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/down.sql @@ -0,0 +1,2 @@ +ALTER TABLE safekeepers ALTER COLUMN scheduling_policy SET DEFAULT 'disabled'; +UPDATE safekeepers SET scheduling_policy = 'disabled' WHERE scheduling_policy = 'pause'; diff --git a/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/up.sql b/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/up.sql new file mode 100644 index 0000000000..9ff75444f3 --- /dev/null +++ b/storage_controller/migrations/2025-01-15-181207_safekeepers_disabled_to_pause/up.sql @@ -0,0 +1,2 @@ +ALTER TABLE safekeepers ALTER COLUMN scheduling_policy SET DEFAULT 'pause'; +UPDATE safekeepers SET scheduling_policy = 'pause' WHERE scheduling_policy = 'disabled'; diff --git a/test_runner/regress/test_storage_controller.py b/test_runner/regress/test_storage_controller.py index b1e1fd81d6..ff479e8fe2 100644 --- a/test_runner/regress/test_storage_controller.py +++ b/test_runner/regress/test_storage_controller.py @@ -3211,7 +3211,7 @@ def test_safekeeper_deployment_time_update(neon_env_builder: NeonEnvBuilder): # some small tests for the scheduling policy querying and returning APIs newest_info = target.get_safekeeper(inserted["id"]) assert newest_info - assert newest_info["scheduling_policy"] == "Disabled" + assert newest_info["scheduling_policy"] == "Pause" target.safekeeper_scheduling_policy(inserted["id"], "Decomissioned") newest_info = target.get_safekeeper(inserted["id"]) assert newest_info From cccc1968487c1f29c6ed190a538178b7c79410ce Mon Sep 17 00:00:00 2001 From: "Alex Chi Z." <4198311+skyzh@users.noreply.github.com> Date: Thu, 16 Jan 2025 10:33:37 -0500 Subject: [PATCH 31/31] refactor(pageserver): make partitioning an ArcSwap (#10377) ## Problem gc-compaction needs the partitioning data to decide the job split. This refactor allows concurrent access/computing the partitioning. ## Summary of changes Make `partitioning` an ArcSwap so that others can access the partitioning while we compute it. Fully eliminate the `repartition is called concurrently` warning when gc-compaction is going on. --------- Signed-off-by: Alex Chi Z --- libs/utils/src/guard_arc_swap.rs | 54 ++++++++++++++++++++ libs/utils/src/lib.rs | 2 + pageserver/src/tenant/timeline.rs | 26 +++++----- pageserver/src/tenant/timeline/compaction.rs | 7 +-- 4 files changed, 70 insertions(+), 19 deletions(-) create mode 100644 libs/utils/src/guard_arc_swap.rs diff --git a/libs/utils/src/guard_arc_swap.rs b/libs/utils/src/guard_arc_swap.rs new file mode 100644 index 0000000000..cec5202460 --- /dev/null +++ b/libs/utils/src/guard_arc_swap.rs @@ -0,0 +1,54 @@ +//! A wrapper around `ArcSwap` that ensures there is only one writer at a time and writes +//! don't block reads. + +use arc_swap::ArcSwap; +use std::sync::Arc; +use tokio::sync::TryLockError; + +pub struct GuardArcSwap { + inner: ArcSwap, + guard: tokio::sync::Mutex<()>, +} + +pub struct Guard<'a, T> { + _guard: tokio::sync::MutexGuard<'a, ()>, + inner: &'a ArcSwap, +} + +impl GuardArcSwap { + pub fn new(inner: T) -> Self { + Self { + inner: ArcSwap::new(Arc::new(inner)), + guard: tokio::sync::Mutex::new(()), + } + } + + pub fn read(&self) -> Arc { + self.inner.load_full() + } + + pub async fn write_guard(&self) -> Guard<'_, T> { + Guard { + _guard: self.guard.lock().await, + inner: &self.inner, + } + } + + pub fn try_write_guard(&self) -> Result, TryLockError> { + let guard = self.guard.try_lock()?; + Ok(Guard { + _guard: guard, + inner: &self.inner, + }) + } +} + +impl Guard<'_, T> { + pub fn read(&self) -> Arc { + self.inner.load_full() + } + + pub fn write(&mut self, value: T) { + self.inner.store(Arc::new(value)); + } +} diff --git a/libs/utils/src/lib.rs b/libs/utils/src/lib.rs index 2c56dd750f..1fb18e9e9a 100644 --- a/libs/utils/src/lib.rs +++ b/libs/utils/src/lib.rs @@ -98,6 +98,8 @@ pub mod try_rcu; pub mod pprof; +pub mod guard_arc_swap; + // Re-export used in macro. Avoids adding git-version as dep in target crates. #[doc(hidden)] pub use git_version; diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs index d6ae11e67d..f24611e1d8 100644 --- a/pageserver/src/tenant/timeline.rs +++ b/pageserver/src/tenant/timeline.rs @@ -51,7 +51,9 @@ use tokio::{ use tokio_util::sync::CancellationToken; use tracing::*; use utils::{ - fs_ext, pausable_failpoint, + fs_ext, + guard_arc_swap::GuardArcSwap, + pausable_failpoint, postgres_client::PostgresClientProtocol, sync::gate::{Gate, GateGuard}, }; @@ -353,8 +355,8 @@ pub struct Timeline { // though let's keep them both for better error visibility. pub initdb_lsn: Lsn, - /// When did we last calculate the partitioning? Make it pub to test cases. - pub(super) partitioning: tokio::sync::Mutex<((KeyPartitioning, SparseKeyPartitioning), Lsn)>, + /// The repartitioning result. Allows a single writer and multiple readers. + pub(crate) partitioning: GuardArcSwap<((KeyPartitioning, SparseKeyPartitioning), Lsn)>, /// Configuration: how often should the partitioning be recalculated. repartition_threshold: u64, @@ -2340,7 +2342,8 @@ impl Timeline { // initial logical size is 0. LogicalSize::empty_initial() }, - partitioning: tokio::sync::Mutex::new(( + + partitioning: GuardArcSwap::new(( (KeyPartitioning::new(), KeyPartitioning::new().into_sparse()), Lsn(0), )), @@ -4028,18 +4031,15 @@ impl Timeline { flags: EnumSet, ctx: &RequestContext, ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> { - let Ok(mut partitioning_guard) = self.partitioning.try_lock() else { + let Ok(mut guard) = self.partitioning.try_write_guard() else { // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline. // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()` // and hence before the compaction task starts. - // Note that there are a third "caller" that will take the `partitioning` lock. It is `gc_compaction_split_jobs` for - // gc-compaction where it uses the repartition data to determine the split jobs. In the future, it might use its own - // heuristics, but for now, we should allow concurrent access to it and let the caller retry compaction. return Err(CompactionError::Other(anyhow!( - "repartition() called concurrently, this is rare and a retry should be fine" + "repartition() called concurrently" ))); }; - let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard; + let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read(); if lsn < *partition_lsn { return Err(CompactionError::Other(anyhow!( "repartition() called with LSN going backwards, this should not happen" @@ -4067,9 +4067,9 @@ impl Timeline { let sparse_partitioning = SparseKeyPartitioning { parts: vec![sparse_ks], }; // no partitioning for metadata keys for now - *partitioning_guard = ((dense_partitioning, sparse_partitioning), lsn); - - Ok((partitioning_guard.0.clone(), partitioning_guard.1)) + let result = ((dense_partitioning, sparse_partitioning), lsn); + guard.write(result.clone()); + Ok(result) } // Is it time to create a new image layer for the given partition? diff --git a/pageserver/src/tenant/timeline/compaction.rs b/pageserver/src/tenant/timeline/compaction.rs index 2042a18e96..06a21f6b3c 100644 --- a/pageserver/src/tenant/timeline/compaction.rs +++ b/pageserver/src/tenant/timeline/compaction.rs @@ -2146,12 +2146,7 @@ impl Timeline { let mut compact_jobs = Vec::new(); // For now, we simply use the key partitioning information; we should do a more fine-grained partitioning // by estimating the amount of files read for a compaction job. We should also partition on LSN. - let ((dense_ks, sparse_ks), _) = { - let Ok(partition) = self.partitioning.try_lock() else { - bail!("failed to acquire partition lock during gc-compaction"); - }; - partition.clone() - }; + let ((dense_ks, sparse_ks), _) = self.partitioning.read().as_ref().clone(); // Truncate the key range to be within user specified compaction range. fn truncate_to( source_start: &Key,