diff --git a/pageserver/src/tenant/remote_timeline_client.rs b/pageserver/src/tenant/remote_timeline_client.rs index 94f42c7827..b910a40547 100644 --- a/pageserver/src/tenant/remote_timeline_client.rs +++ b/pageserver/src/tenant/remote_timeline_client.rs @@ -197,6 +197,7 @@ use utils::backoff::{ self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS, }; use utils::pausable_failpoint; +use utils::shard::ShardNumber; use std::collections::{HashMap, VecDeque}; use std::sync::atomic::{AtomicU32, Ordering}; @@ -2231,6 +2232,28 @@ impl RemoteTimelineClient { UploadQueue::Initialized(x) => x.no_pending_work(), } } + + /// 'foreign' in the sense that it does not belong to this tenant shard. This method + /// is used during GC for other shards to get the index of shard zero. + pub(crate) async fn download_foreign_index( + &self, + shard_number: ShardNumber, + cancel: &CancellationToken, + ) -> Result<(IndexPart, Generation, std::time::SystemTime), DownloadError> { + let foreign_shard_id = TenantShardId { + shard_number, + shard_count: self.tenant_shard_id.shard_count, + tenant_id: self.tenant_shard_id.tenant_id, + }; + download_index_part( + &self.storage_impl, + &foreign_shard_id, + &self.timeline_id, + Generation::MAX, + cancel, + ) + .await + } } pub(crate) struct UploadQueueAccessor<'a> { diff --git a/pageserver/src/tenant/timeline.rs b/pageserver/src/tenant/timeline.rs index 2bc14ec317..a7fadb553f 100644 --- a/pageserver/src/tenant/timeline.rs +++ b/pageserver/src/tenant/timeline.rs @@ -38,6 +38,7 @@ use pageserver_api::{ shard::{ShardIdentity, ShardNumber, TenantShardId}, }; use rand::Rng; +use remote_storage::DownloadError; use serde_with::serde_as; use storage_broker::BrokerClientChannel; use tokio::{ @@ -4774,6 +4775,86 @@ impl Timeline { Ok(()) } + async fn find_gc_time_cutoff( + &self, + pitr: Duration, + cancel: &CancellationToken, + ctx: &RequestContext, + ) -> Result, PageReconstructError> { + debug_assert_current_span_has_tenant_and_timeline_id(); + if self.shard_identity.is_shard_zero() { + // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself + let now = SystemTime::now(); + let time_range = if pitr == Duration::ZERO { + humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid") + } else { + pitr + }; + + // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case) + let time_cutoff = now.checked_sub(time_range).unwrap_or(now); + let timestamp = to_pg_timestamp(time_cutoff); + + let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? { + LsnForTimestamp::Present(lsn) => Some(lsn), + LsnForTimestamp::Future(lsn) => { + // The timestamp is in the future. That sounds impossible, + // but what it really means is that there hasn't been + // any commits since the cutoff timestamp. + // + // In this case we should use the LSN of the most recent commit, + // which is implicitly the last LSN in the log. + debug!("future({})", lsn); + Some(self.get_last_record_lsn()) + } + LsnForTimestamp::Past(lsn) => { + debug!("past({})", lsn); + None + } + LsnForTimestamp::NoData(lsn) => { + debug!("nodata({})", lsn); + None + } + }; + Ok(time_cutoff) + } else { + // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff + // from shard zero's index. The index doesn't explicitly tell us the time cutoff, but we may assume that + // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a + // space cutoff that we would also have respected ourselves. + match self + .remote_client + .download_foreign_index(ShardNumber(0), cancel) + .await + { + Ok((index_part, index_generation, _index_mtime)) => { + tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}", + index_part.metadata.latest_gc_cutoff_lsn()); + Ok(Some(index_part.metadata.latest_gc_cutoff_lsn())) + } + Err(DownloadError::NotFound) => { + // This is unexpected, because during timeline creations shard zero persists to remote + // storage before other shards are called, and during timeline deletion non-zeroth shards are + // deleted before the zeroth one. However, it should be harmless: if we somehow end up in this + // state, then shard zero should _eventually_ write an index when it GCs. + tracing::warn!("GC couldn't find shard zero's index for timeline"); + Ok(None) + } + Err(e) => { + // TODO: this function should return a different error type than page reconstruct error + Err(PageReconstructError::Other(anyhow::anyhow!(e))) + } + } + + // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage + // controller. Otherwise, it is possible that we see the GC cutoff go backwards while shard zero + // is going through a migration if we read the old location's index and it has GC'd ahead of the + // new location. This is legal in principle, but problematic in practice because it might result + // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards + // because they have GC'd past the branch point. + } + } + /// Find the Lsns above which layer files need to be retained on /// garbage collection. /// @@ -4816,40 +4897,7 @@ impl Timeline { // - if PITR interval is set, then this is our cutoff. // - if PITR interval is not set, then we do a lookup // based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases. - let time_cutoff = { - let now = SystemTime::now(); - let time_range = if pitr == Duration::ZERO { - humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid") - } else { - pitr - }; - - // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case) - let time_cutoff = now.checked_sub(time_range).unwrap_or(now); - let timestamp = to_pg_timestamp(time_cutoff); - - match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? { - LsnForTimestamp::Present(lsn) => Some(lsn), - LsnForTimestamp::Future(lsn) => { - // The timestamp is in the future. That sounds impossible, - // but what it really means is that there hasn't been - // any commits since the cutoff timestamp. - // - // In this case we should use the LSN of the most recent commit, - // which is implicitly the last LSN in the log. - debug!("future({})", lsn); - Some(self.get_last_record_lsn()) - } - LsnForTimestamp::Past(lsn) => { - debug!("past({})", lsn); - None - } - LsnForTimestamp::NoData(lsn) => { - debug!("nodata({})", lsn); - None - } - } - }; + let time_cutoff = self.find_gc_time_cutoff(pitr, cancel, ctx).await?; Ok(match (pitr, time_cutoff) { (Duration::ZERO, Some(time_cutoff)) => {