Compare commits

..

12 Commits

Author SHA1 Message Date
Bojan Serafimov
a4bd82cb81 Merge branch 'layer-stats' of github.com:neondatabase/neon into layer-stats 2023-06-10 10:47:42 -04:00
Bojan Serafimov
c64798956d Address comments 2023-06-10 10:46:28 -04:00
bojanserafimov
39bbaecb03 accept suggestion
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
2023-06-10 10:35:13 -04:00
Bojan Serafimov
26dca374eb Add layer stats cli 2023-06-02 16:46:48 -04:00
Heikki Linnakangas
9787227c35 Shield HTTP request handlers from async cancellations. (#4314)
We now spawn a new task for every HTTP request, and wait on the
JoinHandle. If Hyper drops the Future, the spawned task will keep
running. This protects the rest of the pageserver code from unexpected
async cancellations.

This creates a CancellationToken for each request and passes it to the
handler function. If the HTTP request is dropped by the client, the
CancellationToken is signaled. None of the handler functions make use
for the CancellationToken currently, but they now they could.

The CancellationToken arguments also work like documentation. When
you're looking at a function signature and you see that it takes a
CancellationToken as argument, it's a nice hint that the function might
run for a long time, and won't be async cancelled. The default
assumption in the pageserver is now that async functions are not
cancellation-safe anyway, unless explictly marked as such, but this is a
nice extra reminder.

Spawning a task for each request is OK from a performance point of view
because spawning is very cheap in Tokio, and none of our HTTP requests
are very performance critical anyway.

Fixes issue #3478
2023-06-02 08:28:13 -04:00
Joonas Koivunen
ef80a902c8 pg_sni_router: add session_id to more messages (#4403)
See superceded #4390.

- capture log in test
- expand the span to cover init and error reporting
- remove obvious logging by logging only unexpected
2023-06-02 14:59:10 +03:00
Alex Chi Z
66cdba990a refactor: use PersistentLayerDesc for persistent layers (#4398)
## Problem

Part of https://github.com/neondatabase/neon/issues/4373

## Summary of changes

This PR adds `PersistentLayerDesc`, which will be used in LayerMap
mapping and probably layer cache. After this PR and after we change
LayerMap to map to layer desc, we can safely drop RemoteLayerDesc.

---------

Signed-off-by: Alex Chi <iskyzh@gmail.com>
Co-authored-by: bojanserafimov <bojan.serafimov7@gmail.com>
2023-06-01 22:06:28 +03:00
Alex Chi Z
82484e8241 pgserver: add more metrics for better observability (#4323)
## Problem

This PR includes doc changes to the current metrics as well as adding
new metrics. With the new set of metrics, we can quantitatively analyze
the read amp., write amp. and space amp. in the system, when used
together with https://github.com/neondatabase/neonbench

close https://github.com/neondatabase/neon/issues/4312
ref https://github.com/neondatabase/neon/issues/4347

compaction metrics TBD, a novel idea is to print L0 file number and
number of layers in the system, and we can do this in the future when we
start working on compaction.

## Summary of changes

* Add `READ_NUM_FS_LAYERS` for computing read amp.
* Add `MATERIALIZED_PAGE_CACHE_HIT_UPON_REQUEST`.
* Add `GET_RECONSTRUCT_DATA_TIME`. GET_RECONSTRUCT_DATA_TIME +
RECONSTRUCT_TIME + WAIT_LSN_TIME should be approximately total time of
reads.
* Add `5.0` and `10.0` to `STORAGE_IO_TIME_BUCKETS` given some fsync
runs slow (i.e., > 1s) in some cases.
* Some `WAL_REDO` metrics are only used when Postgres is involved in the
redo process.

---------

Signed-off-by: Alex Chi <iskyzh@gmail.com>
2023-06-01 21:46:04 +03:00
Joonas Koivunen
36fee50f4d compute_ctl: enable tracing panic hook (#4375)
compute_ctl can panic, but `tracing` is used for logging. panic stderr
output can interleave with messages from normal logging. The fix is to
use the established way (pageserver, safekeeper, storage_broker) of using
`tracing` to report panics.
2023-06-01 20:12:07 +03:00
bojanserafimov
330083638f Fix stale and misleading comment in LayerMap (#4297) 2023-06-01 05:04:46 +03:00
Konstantin Knizhnik
952d6e43a2 Add pageserver parameter forced_image_creation_limit which can be used… (#4353)
This parameter can be use to restrict number of image layers generated
because of GC request (wanted image layers).
Been set to zero it completely eliminates creation of such image layers.
So it allows to avoid extra storage consumption after merging #3673

## Problem
PR #3673 forces generation of missed image layers. So i short term is
cause cause increase (in worst case up to two times) size of storage.
It was intended (by me) that GC period is comparable with PiTR interval.
But looks like it is not the case now - GC is performed much more
frequently. It may cause the problem with space exhaustion: GC forces
new image creation while large PiTR still prevent GC from collecting old
layers.

## Summary of changes

Add new pageserver parameter` forced_image_creation_limit` which
restrict number of created image layers which are requested by GC.

## Checklist before requesting a review

- [ ] I have performed a self-review of my code.
- [ ] If it is a core feature, I have added thorough tests.
- [ ] Do we need to implement analytics? if so did you add the relevant
metrics to the dashboard?
- [ ] If this PR requires public announcement, mark it with
/release-notes label and add several sentences in this section.

## Checklist before merging

- [ ] Do not forget to reformat commit message to not include the above
checklist
2023-05-31 21:37:20 +03:00
bojanserafimov
b6447462dc Fix layer map correctness bug (#4342) 2023-05-31 12:23:00 -04:00
32 changed files with 954 additions and 510 deletions

1
Cargo.lock generated
View File

@@ -2530,6 +2530,7 @@ dependencies = [
"bytes",
"clap 4.3.0",
"git-version",
"itertools",
"pageserver",
"postgres_ffi",
"svg_fmt",

View File

@@ -33,5 +33,7 @@ pub fn init_tracing_and_logging(default_log_level: &str) -> anyhow::Result<()> {
.init();
tracing::info!("logging and tracing started");
utils::logging::replace_panic_hook_with_tracing_panic_hook().forget();
Ok(())
}

View File

@@ -369,6 +369,11 @@ impl PageServerNode {
evictions_low_residence_duration_metric_threshold: settings
.remove("evictions_low_residence_duration_metric_threshold")
.map(|x| x.to_string()),
gc_feedback: settings
.remove("gc_feedback")
.map(|x| x.parse::<bool>())
.transpose()
.context("Failed to parse 'gc_feedback' as bool")?,
};
// If tenant ID was not specified, generate one
@@ -463,6 +468,11 @@ impl PageServerNode {
evictions_low_residence_duration_metric_threshold: settings
.remove("evictions_low_residence_duration_metric_threshold")
.map(|x| x.to_string()),
gc_feedback: settings
.remove("gc_feedback")
.map(|x| x.parse::<bool>())
.transpose()
.context("Failed to parse 'gc_feedback' as bool")?,
}
};

View File

@@ -52,9 +52,7 @@ completion, or shield the rest of the code from surprise cancellations
by spawning a separate task. The code that handles incoming HTTP
requests, for example, spawns a separate task for each request,
because Hyper will drop the request-handling Future if the HTTP
connection is lost. (FIXME: our HTTP handlers do not do that
currently, but we should fix that. See [issue
3478](https://github.com/neondatabase/neon/issues/3478)).
connection is lost.
#### How to cancel, then?

View File

@@ -223,6 +223,7 @@ pub struct TenantConfig {
pub eviction_policy: Option<serde_json::Value>,
pub min_resident_size_override: Option<u64>,
pub evictions_low_residence_duration_metric_threshold: Option<String>,
pub gc_feedback: Option<bool>,
}
#[serde_as]
@@ -281,6 +282,7 @@ impl TenantConfigRequest {
eviction_policy: None,
min_resident_size_override: None,
evictions_low_residence_duration_metric_threshold: None,
gc_feedback: None,
};
TenantConfigRequest { tenant_id, config }
}

View File

@@ -40,6 +40,12 @@ struct RequestId(String);
///
/// This also handles errors, logging them and converting them to an HTTP error response.
///
/// NB: If the client disconnects, Hyper will drop the Future, without polling it to
/// completion. In other words, the handler must be async cancellation safe! request_span
/// prints a warning to the log when that happens, so that you have some trace of it in
/// the log.
///
///
/// There could be other ways to implement similar functionality:
///
/// * procmacros placed on top of all handler methods

View File

@@ -16,3 +16,4 @@ postgres_ffi.workspace = true
utils.workspace = true
svg_fmt.workspace = true
workspace_hack.workspace = true
itertools.workspace = true

View File

@@ -41,9 +41,18 @@ pub(crate) enum LayerCmd {
/// The id from list-layer command
id: usize,
},
/// Output layer statistics
GetStats {
path: PathBuf,
tenant: String,
timeline: String,
/// The id from list-layer command
id: usize,
},
}
fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
// Return (key, value.len) for all keys, sorted by key.
fn read_delta_file(path: impl AsRef<Path>) -> Result<Vec<(Key, usize)>> {
use pageserver::tenant::blob_io::BlobCursor;
use pageserver::tenant::block_io::BlockReader;
@@ -70,11 +79,48 @@ fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
},
)?;
let mut cursor = BlockCursor::new(&file);
let mut result = vec![];
for (k, v) in all {
let value = cursor.read_blob(v.pos())?;
println!("key:{} value_len:{}", k, value.len());
result.push((k, value.len()));
}
// TODO(chi): special handling for last key?
Ok(result)
}
// We divide the entire i128 keyspace into pre-assigned fixed segments,
// 8MB each. Group keys by segment, and report segment size for each.
//
// 8MB is chosen as the segment size because we're unlikely to make
// s3 partial downloads smaller than 8MB (due to cost). So summarizing
// layer metadata in 8MB segments could be enough to generate good test
// data for write amplification tests.
//
// Note that the segments are fixed, and don't depend on what keyspace
// is actually in use.
fn read_delta_segments(path: impl AsRef<Path>) -> Result<Vec<(i128, usize)>> {
fn key_to_segment(key: &Key) -> i128 {
// A page is 8KB. So 1024 pages are 8MB.
key.to_i128() >> 10
}
use itertools::Itertools;
let delta_metadata = read_delta_file(path)?;
let group_iter = delta_metadata.iter().group_by(|(k, _)| key_to_segment(k));
let group_sizes = group_iter.into_iter().map(|(segment, lengths_group)| {
let sum: usize = lengths_group.map(|(_k, len)| len).sum();
(segment, sum)
});
Ok(group_sizes.collect())
}
fn summarize_delta_file(path: impl AsRef<Path>) -> Result<()> {
// TODO write in some compressed binary format
for (segment, size) in read_delta_segments(path)? {
println!("segment:{} size:{}", segment, size);
}
Ok(())
}
@@ -153,7 +199,38 @@ pub(crate) fn main(cmd: &LayerCmd) -> Result<()> {
);
if layer_file.is_delta {
read_delta_file(layer.path())?;
for (k, len) in read_delta_file(layer.path())? {
println!("key:{} value_len:{}", k, len);
}
} else {
anyhow::bail!("not supported yet :(");
}
break;
}
idx += 1;
}
}
}
LayerCmd::GetStats {
path,
tenant,
timeline,
id,
} => {
let timeline_path = path
.join("tenants")
.join(tenant)
.join("timelines")
.join(timeline);
let mut idx = 0;
for layer in fs::read_dir(timeline_path)? {
let layer = layer?;
if let Some(layer_file) = parse_filename(&layer.file_name().into_string().unwrap())
{
if *id == idx {
if layer_file.is_delta {
summarize_delta_file(layer.path())?;
} else {
anyhow::bail!("not supported yet :(");
}

View File

@@ -108,7 +108,7 @@ pub mod defaults {
#min_resident_size_override = .. # in bytes
#evictions_low_residence_duration_metric_threshold = '{DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD}'
#gc_feedback = false
# [remote_storage]
"###
@@ -828,6 +828,14 @@ impl PageServerConf {
)?);
}
if let Some(gc_feedback) = item.get("gc_feedback") {
t_conf.gc_feedback = Some(
gc_feedback
.as_bool()
.with_context(|| "configure option gc_feedback is not a bool".to_string())?,
);
}
Ok(t_conf)
}

View File

@@ -60,7 +60,7 @@ use utils::serde_percent::Percent;
use crate::{
config::PageServerConf,
task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
tenant::{self, storage_layer::{PersistentLayer, RemoteLayerDesc}, Timeline},
tenant::{self, storage_layer::PersistentLayer, Timeline},
};
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -329,7 +329,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
// If we get far enough in the list that we start to evict layers that are below
// the tenant's min-resident-size threshold, print a warning, and memorize the disk
// usage at that point, in 'usage_planned_min_resident_size_respecting'.
let mut batched: HashMap<_, Vec<Arc<RemoteLayerDesc>>> = HashMap::new();
let mut batched: HashMap<_, Vec<Arc<dyn PersistentLayer>>> = HashMap::new();
let mut warned = None;
let mut usage_planned = usage_pre;
for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
@@ -434,7 +434,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
#[derive(Clone)]
struct EvictionCandidate {
timeline: Arc<Timeline>,
layer: Arc<RemoteLayerDesc>,
layer: Arc<dyn PersistentLayer>,
last_activity_ts: SystemTime,
}

View File

@@ -1,3 +1,6 @@
//!
//! Management HTTP API
//!
use std::collections::HashMap;
use std::sync::Arc;
@@ -46,7 +49,6 @@ use utils::{
};
// Imports only used for testing APIs
#[cfg(feature = "testing")]
use super::models::ConfigureFailpointsRequest;
struct State {
@@ -290,13 +292,19 @@ fn build_timeline_info_common(
}
// healthcheck handler
async fn status_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn status_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
check_permission(&request, None)?;
let config = get_config(&request);
json_response(StatusCode::OK, StatusResponse { id: config.id })
}
async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_create_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let request_data: TimelineCreateRequest = json_request(&mut request).await?;
check_permission(&request, Some(tenant_id))?;
@@ -332,7 +340,10 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
.await
}
async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_list_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let include_non_incremental_logical_size: Option<bool> =
parse_query_param(&request, "include-non-incremental-logical-size")?;
@@ -366,7 +377,10 @@ async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>,
json_response(StatusCode::OK, response_data)
}
async fn timeline_detail_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_detail_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
let include_non_incremental_logical_size: Option<bool> =
@@ -400,7 +414,10 @@ async fn timeline_detail_handler(request: Request<Body>) -> Result<Response<Body
json_response(StatusCode::OK, timeline_info)
}
async fn get_lsn_by_timestamp_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn get_lsn_by_timestamp_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -424,7 +441,10 @@ async fn get_lsn_by_timestamp_handler(request: Request<Body>) -> Result<Response
json_response(StatusCode::OK, result)
}
async fn tenant_attach_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_attach_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -460,7 +480,10 @@ async fn tenant_attach_handler(mut request: Request<Body>) -> Result<Response<Bo
json_response(StatusCode::ACCEPTED, ())
}
async fn timeline_delete_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_delete_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -474,7 +497,10 @@ async fn timeline_delete_handler(request: Request<Body>) -> Result<Response<Body
json_response(StatusCode::OK, ())
}
async fn tenant_detach_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_detach_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
let detach_ignored: Option<bool> = parse_query_param(&request, "detach_ignored")?;
@@ -488,7 +514,10 @@ async fn tenant_detach_handler(request: Request<Body>) -> Result<Response<Body>,
json_response(StatusCode::OK, ())
}
async fn tenant_load_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_load_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -508,7 +537,10 @@ async fn tenant_load_handler(request: Request<Body>) -> Result<Response<Body>, A
json_response(StatusCode::ACCEPTED, ())
}
async fn tenant_ignore_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_ignore_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -521,7 +553,10 @@ async fn tenant_ignore_handler(request: Request<Body>) -> Result<Response<Body>,
json_response(StatusCode::OK, ())
}
async fn tenant_list_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_list_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
check_permission(&request, None)?;
let response_data = mgr::list_tenants()
@@ -541,7 +576,10 @@ async fn tenant_list_handler(request: Request<Body>) -> Result<Response<Body>, A
json_response(StatusCode::OK, response_data)
}
async fn tenant_status(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_status(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -581,7 +619,10 @@ async fn tenant_status(request: Request<Body>) -> Result<Response<Body>, ApiErro
/// Note: we don't update the cached size and prometheus metric here.
/// The retention period might be different, and it's nice to have a method to just calculate it
/// without modifying anything anyway.
async fn tenant_size_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_size_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
let inputs_only: Option<bool> = parse_query_param(&request, "inputs_only")?;
@@ -646,7 +687,10 @@ async fn tenant_size_handler(request: Request<Body>) -> Result<Response<Body>, A
)
}
async fn layer_map_info_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn layer_map_info_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
let reset: LayerAccessStatsReset =
@@ -660,7 +704,10 @@ async fn layer_map_info_handler(request: Request<Body>) -> Result<Response<Body>
json_response(StatusCode::OK, layer_map_info)
}
async fn layer_download_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn layer_download_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
@@ -683,7 +730,10 @@ async fn layer_download_handler(request: Request<Body>) -> Result<Response<Body>
}
}
async fn evict_timeline_layer_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn evict_timeline_layer_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
@@ -761,7 +811,10 @@ pub fn html_response(status: StatusCode, data: String) -> Result<Response<Body>,
Ok(response)
}
async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn tenant_create_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let request_data: TenantCreateRequest = json_request(&mut request).await?;
let target_tenant_id = request_data.new_tenant_id;
check_permission(&request, None)?;
@@ -808,7 +861,10 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
)
}
async fn get_tenant_config_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn get_tenant_config_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -834,6 +890,7 @@ async fn get_tenant_config_handler(request: Request<Body>) -> Result<Response<Bo
async fn update_tenant_config_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let request_data: TenantConfigRequest = json_request(&mut request).await?;
let tenant_id = request_data.tenant_id;
@@ -851,8 +908,10 @@ async fn update_tenant_config_handler(
}
/// Testing helper to transition a tenant to [`crate::tenant::TenantState::Broken`].
#[cfg(feature = "testing")]
async fn handle_tenant_break(r: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn handle_tenant_break(
r: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&r, "tenant_id")?;
let tenant = crate::tenant::mgr::get_tenant(tenant_id, true)
@@ -864,8 +923,10 @@ async fn handle_tenant_break(r: Request<Body>) -> Result<Response<Body>, ApiErro
json_response(StatusCode::OK, ())
}
#[cfg(feature = "testing")]
async fn failpoints_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn failpoints_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
if !fail::has_failpoints() {
return Err(ApiError::BadRequest(anyhow!(
"Cannot manage failpoints because pageserver was compiled without failpoints support"
@@ -898,7 +959,10 @@ async fn failpoints_handler(mut request: Request<Body>) -> Result<Response<Body>
}
// Run GC immediately on given timeline.
async fn timeline_gc_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_gc_handler(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -917,8 +981,10 @@ async fn timeline_gc_handler(mut request: Request<Body>) -> Result<Response<Body
}
// Run compaction immediately on given timeline.
#[cfg(feature = "testing")]
async fn timeline_compact_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_compact_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -939,8 +1005,10 @@ async fn timeline_compact_handler(request: Request<Body>) -> Result<Response<Bod
}
// Run checkpoint immediately on given timeline.
#[cfg(feature = "testing")]
async fn timeline_checkpoint_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn timeline_checkpoint_handler(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -964,6 +1032,7 @@ async fn timeline_checkpoint_handler(request: Request<Body>) -> Result<Response<
async fn timeline_download_remote_layers_handler_post(
mut request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
@@ -979,6 +1048,7 @@ async fn timeline_download_remote_layers_handler_post(
async fn timeline_download_remote_layers_handler_get(
request: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
check_permission(&request, Some(tenant_id))?;
@@ -1002,7 +1072,10 @@ async fn active_timeline_of_active_tenant(
.map_err(ApiError::NotFound)
}
async fn always_panic_handler(req: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn always_panic_handler(
req: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
// Deliberately cause a panic to exercise the panic hook registered via std::panic::set_hook().
// For pageserver, the relevant panic hook is `tracing_panic_hook` , and the `sentry` crate's wrapper around it.
// Use catch_unwind to ensure that tokio nor hyper are distracted by our panic.
@@ -1013,7 +1086,10 @@ async fn always_panic_handler(req: Request<Body>) -> Result<Response<Body>, ApiE
json_response(StatusCode::NO_CONTENT, ())
}
async fn disk_usage_eviction_run(mut r: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn disk_usage_eviction_run(
mut r: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
check_permission(&r, None)?;
#[derive(Debug, Clone, Copy, serde::Serialize, serde::Deserialize)]
@@ -1103,8 +1179,10 @@ async fn handler_404(_: Request<Body>) -> Result<Response<Body>, ApiError> {
)
}
#[cfg(feature = "testing")]
async fn post_tracing_event_handler(mut r: Request<Body>) -> Result<Response<Body>, ApiError> {
async fn post_tracing_event_handler(
mut r: Request<Body>,
_cancel: CancellationToken,
) -> Result<Response<Body>, ApiError> {
#[derive(Debug, serde::Deserialize)]
#[serde(rename_all = "lowercase")]
enum Level {
@@ -1134,6 +1212,85 @@ async fn post_tracing_event_handler(mut r: Request<Body>) -> Result<Response<Bod
json_response(StatusCode::OK, ())
}
/// Common functionality of all the HTTP API handlers.
///
/// - Adds a tracing span to each request (by `request_span`)
/// - Logs the request depending on the request method (by `request_span`)
/// - Logs the response if it was not successful (by `request_span`
/// - Shields the handler function from async cancellations. Hyper can drop the handler
/// Future if the connection to the client is lost, but most of the pageserver code is
/// not async cancellation safe. This converts the dropped future into a graceful cancellation
/// request with a CancellationToken.
async fn api_handler<R, H>(request: Request<Body>, handler: H) -> Result<Response<Body>, ApiError>
where
R: std::future::Future<Output = Result<Response<Body>, ApiError>> + Send + 'static,
H: FnOnce(Request<Body>, CancellationToken) -> R + Send + Sync + 'static,
{
// Spawn a new task to handle the request, to protect the handler from unexpected
// async cancellations. Most pageserver functions are not async cancellation safe.
// We arm a drop-guard, so that if Hyper drops the Future, we signal the task
// with the cancellation token.
let token = CancellationToken::new();
let cancel_guard = token.clone().drop_guard();
let result = request_span(request, move |r| async {
let handle = tokio::spawn(
async {
let token_cloned = token.clone();
let result = handler(r, token).await;
if token_cloned.is_cancelled() {
info!("Cancelled request finished");
}
result
}
.in_current_span(),
);
match handle.await {
Ok(result) => result,
Err(e) => {
// The handler task panicked. We have a global panic handler that logs the
// panic with its backtrace, so no need to log that here. Only log a brief
// message to make it clear that we returned the error to the client.
error!("HTTP request handler task panicked: {e:#}");
// Don't return an Error here, because then fallback error handler that was
// installed in make_router() will print the error. Instead, construct the
// HTTP error response and return that.
Ok(
ApiError::InternalServerError(anyhow!("HTTP request handler task panicked"))
.into_response(),
)
}
}
})
.await;
cancel_guard.disarm();
result
}
/// Like api_handler, but returns an error response if the server is built without
/// the 'testing' feature.
async fn testing_api_handler<R, H>(
desc: &str,
request: Request<Body>,
handler: H,
) -> Result<Response<Body>, ApiError>
where
R: std::future::Future<Output = Result<Response<Body>, ApiError>> + Send + 'static,
H: FnOnce(Request<Body>, CancellationToken) -> R + Send + Sync + 'static,
{
if cfg!(feature = "testing") {
api_handler(request, handler).await
} else {
std::future::ready(Err(ApiError::BadRequest(anyhow!(
"Cannot {desc} because pageserver was compiled without testing APIs",
))))
.await
}
}
pub fn make_router(
conf: &'static PageServerConf,
launch_ts: &'static LaunchTimestamp,
@@ -1163,26 +1320,6 @@ pub fn make_router(
.expect("construct launch timestamp header middleware"),
);
macro_rules! testing_api {
($handler_desc:literal, $handler:path $(,)?) => {{
#[cfg(not(feature = "testing"))]
async fn cfg_disabled(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
Err(ApiError::BadRequest(anyhow!(concat!(
"Cannot ",
$handler_desc,
" because pageserver was compiled without testing APIs",
))))
}
#[cfg(feature = "testing")]
let handler = $handler;
#[cfg(not(feature = "testing"))]
let handler = cfg_disabled;
move |r| request_span(r, handler)
}};
}
Ok(router
.data(Arc::new(
State::new(
@@ -1194,92 +1331,88 @@ pub fn make_router(
)
.context("Failed to initialize router state")?,
))
.get("/v1/status", |r| request_span(r, status_handler))
.put(
"/v1/failpoints",
testing_api!("manage failpoints", failpoints_handler),
)
.get("/v1/tenant", |r| request_span(r, tenant_list_handler))
.post("/v1/tenant", |r| request_span(r, tenant_create_handler))
.get("/v1/tenant/:tenant_id", |r| request_span(r, tenant_status))
.get("/v1/status", |r| api_handler(r, status_handler))
.put("/v1/failpoints", |r| {
testing_api_handler("manage failpoints", r, failpoints_handler)
})
.get("/v1/tenant", |r| api_handler(r, tenant_list_handler))
.post("/v1/tenant", |r| api_handler(r, tenant_create_handler))
.get("/v1/tenant/:tenant_id", |r| api_handler(r, tenant_status))
.get("/v1/tenant/:tenant_id/synthetic_size", |r| {
request_span(r, tenant_size_handler)
api_handler(r, tenant_size_handler)
})
.put("/v1/tenant/config", |r| {
request_span(r, update_tenant_config_handler)
api_handler(r, update_tenant_config_handler)
})
.get("/v1/tenant/:tenant_id/config", |r| {
request_span(r, get_tenant_config_handler)
api_handler(r, get_tenant_config_handler)
})
.get("/v1/tenant/:tenant_id/timeline", |r| {
request_span(r, timeline_list_handler)
api_handler(r, timeline_list_handler)
})
.post("/v1/tenant/:tenant_id/timeline", |r| {
request_span(r, timeline_create_handler)
api_handler(r, timeline_create_handler)
})
.post("/v1/tenant/:tenant_id/attach", |r| {
request_span(r, tenant_attach_handler)
api_handler(r, tenant_attach_handler)
})
.post("/v1/tenant/:tenant_id/detach", |r| {
request_span(r, tenant_detach_handler)
api_handler(r, tenant_detach_handler)
})
.post("/v1/tenant/:tenant_id/load", |r| {
request_span(r, tenant_load_handler)
api_handler(r, tenant_load_handler)
})
.post("/v1/tenant/:tenant_id/ignore", |r| {
request_span(r, tenant_ignore_handler)
api_handler(r, tenant_ignore_handler)
})
.get("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
request_span(r, timeline_detail_handler)
api_handler(r, timeline_detail_handler)
})
.get(
"/v1/tenant/:tenant_id/timeline/:timeline_id/get_lsn_by_timestamp",
|r| request_span(r, get_lsn_by_timestamp_handler),
|r| api_handler(r, get_lsn_by_timestamp_handler),
)
.put("/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc", |r| {
request_span(r, timeline_gc_handler)
api_handler(r, timeline_gc_handler)
})
.put("/v1/tenant/:tenant_id/timeline/:timeline_id/compact", |r| {
testing_api_handler("run timeline compaction", r, timeline_compact_handler)
})
.put(
"/v1/tenant/:tenant_id/timeline/:timeline_id/compact",
testing_api!("run timeline compaction", timeline_compact_handler),
)
.put(
"/v1/tenant/:tenant_id/timeline/:timeline_id/checkpoint",
testing_api!("run timeline checkpoint", timeline_checkpoint_handler),
|r| testing_api_handler("run timeline checkpoint", r, timeline_checkpoint_handler),
)
.post(
"/v1/tenant/:tenant_id/timeline/:timeline_id/download_remote_layers",
|r| request_span(r, timeline_download_remote_layers_handler_post),
|r| api_handler(r, timeline_download_remote_layers_handler_post),
)
.get(
"/v1/tenant/:tenant_id/timeline/:timeline_id/download_remote_layers",
|r| request_span(r, timeline_download_remote_layers_handler_get),
|r| api_handler(r, timeline_download_remote_layers_handler_get),
)
.delete("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
request_span(r, timeline_delete_handler)
api_handler(r, timeline_delete_handler)
})
.get("/v1/tenant/:tenant_id/timeline/:timeline_id/layer", |r| {
request_span(r, layer_map_info_handler)
api_handler(r, layer_map_info_handler)
})
.get(
"/v1/tenant/:tenant_id/timeline/:timeline_id/layer/:layer_file_name",
|r| request_span(r, layer_download_handler),
|r| api_handler(r, layer_download_handler),
)
.delete(
"/v1/tenant/:tenant_id/timeline/:timeline_id/layer/:layer_file_name",
|r| request_span(r, evict_timeline_layer_handler),
|r| api_handler(r, evict_timeline_layer_handler),
)
.put("/v1/disk_usage_eviction/run", |r| {
request_span(r, disk_usage_eviction_run)
api_handler(r, disk_usage_eviction_run)
})
.put("/v1/tenant/:tenant_id/break", |r| {
testing_api_handler("set tenant state to broken", r, handle_tenant_break)
})
.get("/v1/panic", |r| api_handler(r, always_panic_handler))
.post("/v1/tracing/event", |r| {
testing_api_handler("emit a tracing event", r, post_tracing_event_handler)
})
.put(
"/v1/tenant/:tenant_id/break",
testing_api!("set tenant state to broken", handle_tenant_break),
)
.get("/v1/panic", |r| request_span(r, always_panic_handler))
.post(
"/v1/tracing/event",
testing_api!("emit a tracing event", post_tracing_event_handler),
)
.any(handler_404))
}

View File

@@ -84,6 +84,16 @@ pub static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
.expect("failed to define a metric")
});
static READ_NUM_FS_LAYERS: Lazy<HistogramVec> = Lazy::new(|| {
register_histogram_vec!(
"pageserver_read_num_fs_layers",
"Number of persistent layers accessed for processing a read request, including those in the cache",
&["tenant_id", "timeline_id"],
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 10.0, 20.0, 50.0, 100.0],
)
.expect("failed to define a metric")
});
// Metrics collected on operations on the storage repository.
static RECONSTRUCT_TIME: Lazy<HistogramVec> = Lazy::new(|| {
register_histogram_vec!(
@@ -95,6 +105,25 @@ static RECONSTRUCT_TIME: Lazy<HistogramVec> = Lazy::new(|| {
.expect("failed to define a metric")
});
static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounterVec> = Lazy::new(|| {
register_int_counter_vec!(
"pageserver_materialized_cache_hits_direct_total",
"Number of cache hits from materialized page cache without redo",
&["tenant_id", "timeline_id"]
)
.expect("failed to define a metric")
});
static GET_RECONSTRUCT_DATA_TIME: Lazy<HistogramVec> = Lazy::new(|| {
register_histogram_vec!(
"pageserver_getpage_get_reconstruct_data_seconds",
"Time spent in get_reconstruct_value_data",
&["tenant_id", "timeline_id"],
CRITICAL_OP_BUCKETS.into(),
)
.expect("failed to define a metric")
});
static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounterVec> = Lazy::new(|| {
register_int_counter_vec!(
"pageserver_materialized_cache_hits_total",
@@ -354,6 +383,7 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
0.001000, // 1000 usec
0.030, // 30 ms
1.000, // 1000 ms
30.000, // 30000 ms
];
const STORAGE_IO_TIME_OPERATIONS: &[&str] = &[
@@ -622,7 +652,7 @@ pub static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
pub static WAL_REDO_WAIT_TIME: Lazy<Histogram> = Lazy::new(|| {
register_histogram!(
"pageserver_wal_redo_wait_seconds",
"Time spent waiting for access to the WAL redo process",
"Time spent waiting for access to the Postgres WAL redo process",
redo_histogram_time_buckets!(),
)
.expect("failed to define a metric")
@@ -631,7 +661,7 @@ pub static WAL_REDO_WAIT_TIME: Lazy<Histogram> = Lazy::new(|| {
pub static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
register_histogram!(
"pageserver_wal_redo_records_histogram",
"Histogram of number of records replayed per redo",
"Histogram of number of records replayed per redo in the Postgres WAL redo process",
redo_histogram_count_buckets!(),
)
.expect("failed to define a metric")
@@ -640,7 +670,7 @@ pub static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
pub static WAL_REDO_BYTES_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
register_histogram!(
"pageserver_wal_redo_bytes_histogram",
"Histogram of number of records replayed per redo",
"Histogram of number of records replayed per redo sent to Postgres",
redo_bytes_histogram_count_buckets!(),
)
.expect("failed to define a metric")
@@ -723,7 +753,9 @@ pub struct TimelineMetrics {
tenant_id: String,
timeline_id: String,
pub reconstruct_time_histo: Histogram,
pub get_reconstruct_data_time_histo: Histogram,
pub materialized_page_cache_hit_counter: GenericCounter<AtomicU64>,
pub materialized_page_cache_hit_upon_request_counter: GenericCounter<AtomicU64>,
pub flush_time_histo: StorageTimeMetrics,
pub compact_time_histo: StorageTimeMetrics,
pub create_images_time_histo: StorageTimeMetrics,
@@ -734,6 +766,7 @@ pub struct TimelineMetrics {
pub last_record_gauge: IntGauge,
pub wait_lsn_time_histo: Histogram,
pub resident_physical_size_gauge: UIntGauge,
pub read_num_fs_layers: Histogram,
/// copy of LayeredTimeline.current_logical_size
pub current_logical_size_gauge: UIntGauge,
pub num_persistent_files_created: IntCounter,
@@ -753,6 +786,9 @@ impl TimelineMetrics {
let reconstruct_time_histo = RECONSTRUCT_TIME
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
let get_reconstruct_data_time_histo = GET_RECONSTRUCT_DATA_TIME
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
let materialized_page_cache_hit_counter = MATERIALIZED_PAGE_CACHE_HIT
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
@@ -794,6 +830,12 @@ impl TimelineMetrics {
let evictions = EVICTIONS
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
let read_num_fs_layers = READ_NUM_FS_LAYERS
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
let materialized_page_cache_hit_upon_request_counter = MATERIALIZED_PAGE_CACHE_HIT_DIRECT
.get_metric_with_label_values(&[&tenant_id, &timeline_id])
.unwrap();
let evictions_with_low_residence_duration =
evictions_with_low_residence_duration_builder.build(&tenant_id, &timeline_id);
@@ -801,7 +843,9 @@ impl TimelineMetrics {
tenant_id,
timeline_id,
reconstruct_time_histo,
get_reconstruct_data_time_histo,
materialized_page_cache_hit_counter,
materialized_page_cache_hit_upon_request_counter,
flush_time_histo,
compact_time_histo,
create_images_time_histo,
@@ -819,6 +863,7 @@ impl TimelineMetrics {
evictions_with_low_residence_duration: std::sync::RwLock::new(
evictions_with_low_residence_duration,
),
read_num_fs_layers,
}
}
}
@@ -828,7 +873,9 @@ impl Drop for TimelineMetrics {
let tenant_id = &self.tenant_id;
let timeline_id = &self.timeline_id;
let _ = RECONSTRUCT_TIME.remove_label_values(&[tenant_id, timeline_id]);
let _ = GET_RECONSTRUCT_DATA_TIME.remove_label_values(&[tenant_id, timeline_id]);
let _ = MATERIALIZED_PAGE_CACHE_HIT.remove_label_values(&[tenant_id, timeline_id]);
let _ = MATERIALIZED_PAGE_CACHE_HIT_DIRECT.remove_label_values(&[tenant_id, timeline_id]);
let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, timeline_id]);
let _ = WAIT_LSN_TIME.remove_label_values(&[tenant_id, timeline_id]);
let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
@@ -836,6 +883,8 @@ impl Drop for TimelineMetrics {
let _ = NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, timeline_id]);
let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, timeline_id]);
let _ = EVICTIONS.remove_label_values(&[tenant_id, timeline_id]);
let _ = READ_NUM_FS_LAYERS.remove_label_values(&[tenant_id, timeline_id]);
self.evictions_with_low_residence_duration
.write()
.unwrap()

View File

@@ -84,7 +84,6 @@ pub mod block_io;
pub mod disk_btree;
pub(crate) mod ephemeral_file;
pub mod layer_map;
pub mod layer_cache;
pub mod metadata;
mod par_fsync;
@@ -3210,6 +3209,7 @@ pub mod harness {
evictions_low_residence_duration_metric_threshold: Some(
tenant_conf.evictions_low_residence_duration_metric_threshold,
),
gc_feedback: Some(tenant_conf.gc_feedback),
}
}
}

View File

@@ -99,6 +99,7 @@ pub struct TenantConf {
// See the corresponding metric's help string.
#[serde(with = "humantime_serde")]
pub evictions_low_residence_duration_metric_threshold: Duration,
pub gc_feedback: bool,
}
/// Same as TenantConf, but this struct preserves the information about
@@ -175,6 +176,10 @@ pub struct TenantConfOpt {
#[serde(with = "humantime_serde")]
#[serde(default)]
pub evictions_low_residence_duration_metric_threshold: Option<Duration>,
#[serde(skip_serializing_if = "Option::is_none")]
#[serde(default)]
pub gc_feedback: Option<bool>,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -242,6 +247,7 @@ impl TenantConfOpt {
evictions_low_residence_duration_metric_threshold: self
.evictions_low_residence_duration_metric_threshold
.unwrap_or(global_conf.evictions_low_residence_duration_metric_threshold),
gc_feedback: self.gc_feedback.unwrap_or(global_conf.gc_feedback),
}
}
}
@@ -278,6 +284,7 @@ impl Default for TenantConf {
DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD,
)
.expect("cannot parse default evictions_low_residence_duration_metric_threshold"),
gc_feedback: false,
}
}
}
@@ -372,6 +379,7 @@ impl TryFrom<&'_ models::TenantConfig> for TenantConfOpt {
))?,
);
}
tenant_conf.gc_feedback = request_data.gc_feedback;
Ok(tenant_conf)
}

View File

@@ -1,34 +0,0 @@
use std::{
collections::HashMap,
sync::{Arc, Mutex},
};
use super::storage_layer::{LayerFileName, PersistentLayer, RemoteLayerDesc};
pub struct LayerCache {
layers: Mutex<HashMap<LayerFileName, Arc<dyn PersistentLayer>>>,
}
impl LayerCache {
pub fn new() -> Self {
Self {
layers: Mutex::new(HashMap::new()),
}
}
pub fn get(&self, layer_fname: &LayerFileName) -> Option<Arc<dyn PersistentLayer>> {
let guard: std::sync::MutexGuard<HashMap<LayerFileName, Arc<dyn PersistentLayer>>> =
self.layers.lock().unwrap();
guard.get(layer_fname).cloned()
}
pub fn contains(&self, layer_fname: &LayerFileName) -> bool {
let guard = self.layers.lock().unwrap();
guard.contains_key(layer_fname)
}
pub fn insert(&self, layer_fname: LayerFileName, persistent_layer: Arc<dyn PersistentLayer>) {
let mut guard = self.layers.lock().unwrap();
guard.insert(layer_fname, persistent_layer);
}
}

View File

@@ -61,7 +61,6 @@ use historic_layer_coverage::BufferedHistoricLayerCoverage;
pub use historic_layer_coverage::Replacement;
use super::storage_layer::range_eq;
use super::storage_layer::PersistentLayer;
///
/// LayerMap tracks what layers exist on a timeline.
@@ -139,19 +138,24 @@ where
self.layer_map.remove_historic_noflush(layer)
}
/// Ensure the downloaded layer matches existing layer.
/// Replaces existing layer iff it is the `expected`.
///
/// Returned `Replacement` describes succeeding in checking or the reason why it could not
/// If the expected layer has been removed it will not be inserted by this function.
///
/// Returned `Replacement` describes succeeding in replacement or the reason why it could not
/// be done.
pub fn ensure_consistent(
&self,
///
/// TODO replacement can be done without buffering and rebuilding layer map updates.
/// One way to do that is to add a layer of indirection for returned values, so
/// that we can replace values only by updating a hashmap.
pub fn replace_historic(
&mut self,
expected: &Arc<L>,
new: Arc<dyn PersistentLayer>,
new: Arc<L>,
) -> anyhow::Result<Replacement<Arc<L>>> {
fail::fail_point!("layermap-replace-notfound", |_| Ok(Replacement::NotFound));
self.layer_map
.ensure_consistent_noflush(expected, new)
self.layer_map.replace_historic_noflush(expected, new)
}
// We will flush on drop anyway, but this method makes it
@@ -305,16 +309,16 @@ where
}
}
pub(self) fn ensure_consistent_noflush(
&self,
pub(self) fn replace_historic_noflush(
&mut self,
expected: &Arc<L>,
new: Arc<dyn PersistentLayer>,
new: Arc<L>,
) -> anyhow::Result<Replacement<Arc<L>>> {
let key = historic_layer_coverage::LayerKey::from(&**expected);
let other = historic_layer_coverage::LayerKey::from(&*new);
let expected_l0 = Self::is_l0(expected);
let new_l0 = LayerMap::<dyn PersistentLayer>::is_l0(&*new);
let new_l0 = Self::is_l0(&new);
anyhow::ensure!(
key == other,
@@ -341,7 +345,17 @@ where
None
};
Ok(Replacement::Replaced { in_buffered: false })
let replaced = self.historic.replace(&key, new.clone(), |existing| {
Self::compare_arced_layers(existing, expected)
});
if let Replacement::Replaced { .. } = &replaced {
if let Some(index) = l0_index {
self.l0_delta_layers[index] = new;
}
}
Ok(replaced)
}
/// Helper function for BatchedUpdates::drop.

View File

@@ -204,6 +204,35 @@ fn test_off_by_one() {
assert_eq!(version.image_coverage.query(5), None);
}
/// White-box regression test, checking for incorrect removal of node at key.end
#[test]
fn test_regression() {
let mut map = HistoricLayerCoverage::<String>::new();
map.insert(
LayerKey {
key: 0..5,
lsn: 0..5,
is_image: false,
},
"Layer 1".to_string(),
);
map.insert(
LayerKey {
key: 0..5,
lsn: 1..2,
is_image: false,
},
"Layer 2".to_string(),
);
// If an insertion operation improperly deletes the endpoint of a previous layer
// (which is more likely to happen with layers that collide on key.end), we will
// end up with an infinite layer, covering the entire keyspace. Here we assert
// that there's no layer at key 100 because we didn't insert any layer there.
let version = map.get_version(100).unwrap();
assert_eq!(version.delta_coverage.query(100), None);
}
/// Cover edge cases where layers begin or end on the same key
#[test]
fn test_key_collision() {

View File

@@ -1,8 +1,8 @@
use std::ops::Range;
// TODO the `im` crate has 20x more downloads and also has
// persistent/immutable BTree. It also runs a bit faster but
// results are not the same on some tests.
// NOTE the `im` crate has 20x more downloads and also has
// persistent/immutable BTree. But it's bugged so rpds is a
// better choice https://github.com/neondatabase/neon/issues/3395
use rpds::RedBlackTreeMapSync;
/// Data structure that can efficiently:
@@ -10,19 +10,22 @@ use rpds::RedBlackTreeMapSync;
/// - iterate the latest layers in a key range
/// - insert layers in non-decreasing lsn.start order
///
/// The struct is parameterized over Value for easier
/// testing, but in practice it's some sort of layer.
/// For a detailed explanation and justification of this approach, see:
/// https://neon.tech/blog/persistent-structures-in-neons-wal-indexing
///
/// NOTE The struct is parameterized over Value for easier
/// testing, but in practice it's some sort of layer.
pub struct LayerCoverage<Value> {
/// For every change in coverage (as we sweep the key space)
/// we store (lsn.end, value).
///
/// We use an immutable/persistent tree so that we can keep historic
/// versions of this coverage without cloning the whole thing and
/// incurring quadratic memory cost. See HistoricLayerCoverage.
/// NOTE We use an immutable/persistent tree so that we can keep historic
/// versions of this coverage without cloning the whole thing and
/// incurring quadratic memory cost. See HistoricLayerCoverage.
///
/// We use the Sync version of the map because we want Self to
/// be Sync. Using nonsync might be faster, if we can work with
/// that.
/// NOTE We use the Sync version of the map because we want Self to
/// be Sync. Using nonsync might be faster, if we can work with
/// that.
nodes: RedBlackTreeMapSync<i128, Option<(u64, Value)>>,
}
@@ -41,6 +44,13 @@ impl<Value: Clone> LayerCoverage<Value> {
/// Helper function to subdivide the key range without changing any values
///
/// This operation has no semantic effect by itself. It only helps us pin in
/// place the part of the coverage we don't want to change when inserting.
///
/// As an analogy, think of a polygon. If you add a vertex along one of the
/// segments, the polygon is still the same, but it behaves differently when
/// we move or delete one of the other points.
///
/// Complexity: O(log N)
fn add_node(&mut self, key: i128) {
let value = match self.nodes.range(..=key).last() {
@@ -74,7 +84,7 @@ impl<Value: Clone> LayerCoverage<Value> {
let mut to_update = Vec::new();
let mut to_remove = Vec::new();
let mut prev_covered = false;
for (k, node) in self.nodes.range(key.clone()) {
for (k, node) in self.nodes.range(key) {
let needs_cover = match node {
None => true,
Some((h, _)) => h < &lsn.end,
@@ -87,9 +97,8 @@ impl<Value: Clone> LayerCoverage<Value> {
}
prev_covered = needs_cover;
}
if !prev_covered {
to_remove.push(key.end);
}
// TODO check if the nodes inserted at key.start and key.end are safe
// to remove. It's fine to keep them but they could be redundant.
for k in to_update {
self.nodes.insert_mut(k, Some((lsn.end, value.clone())));
}

View File

@@ -779,7 +779,6 @@ pub async fn immediate_gc(
Ok(wait_task_done)
}
#[cfg(feature = "testing")]
pub async fn immediate_compact(
tenant_id: TenantId,
timeline_id: TimelineId,

View File

@@ -4,6 +4,7 @@ pub mod delta_layer;
mod filename;
mod image_layer;
mod inmemory_layer;
mod layer_desc;
mod remote_layer;
use crate::config::PageServerConf;
@@ -37,7 +38,8 @@ pub use delta_layer::{DeltaLayer, DeltaLayerWriter};
pub use filename::{DeltaFileName, ImageFileName, LayerFileName};
pub use image_layer::{ImageLayer, ImageLayerWriter};
pub use inmemory_layer::InMemoryLayer;
pub use remote_layer::RemoteLayerDesc;
pub use layer_desc::PersistentLayerDesc;
pub use remote_layer::RemoteLayer;
use super::layer_map::BatchedUpdates;
@@ -406,14 +408,23 @@ pub type LayerKeyIter<'i> = Box<dyn Iterator<Item = (Key, Lsn, u64)> + 'i>;
/// An image layer is a snapshot of all the data in a key-range, at a single
/// LSN.
pub trait PersistentLayer: Layer {
fn get_tenant_id(&self) -> TenantId;
/// Get the layer descriptor.
fn layer_desc(&self) -> &PersistentLayerDesc;
fn get_tenant_id(&self) -> TenantId {
self.layer_desc().tenant_id
}
/// Identify the timeline this layer belongs to
fn get_timeline_id(&self) -> TimelineId;
fn get_timeline_id(&self) -> TimelineId {
self.layer_desc().timeline_id
}
/// File name used for this layer, both in the pageserver's local filesystem
/// state as well as in the remote storage.
fn filename(&self) -> LayerFileName;
fn filename(&self) -> LayerFileName {
self.layer_desc().filename()
}
// Path to the layer file in the local filesystem.
// `None` for `RemoteLayer`.
@@ -431,6 +442,14 @@ pub trait PersistentLayer: Layer {
/// Permanently remove this layer from disk.
fn delete_resident_layer_file(&self) -> Result<()>;
fn downcast_remote_layer(self: Arc<Self>) -> Option<std::sync::Arc<RemoteLayer>> {
None
}
fn is_remote_layer(&self) -> bool {
false
}
/// Returns None if the layer file size is not known.
///
/// Should not change over the lifetime of the layer object because
@@ -442,6 +461,16 @@ pub trait PersistentLayer: Layer {
fn access_stats(&self) -> &LayerAccessStats;
}
pub fn downcast_remote_layer(
layer: &Arc<dyn PersistentLayer>,
) -> Option<std::sync::Arc<RemoteLayer>> {
if layer.is_remote_layer() {
Arc::clone(layer).downcast_remote_layer()
} else {
None
}
}
/// Holds metadata about a layer without any content. Used mostly for testing.
///
/// To use filenames as fixtures, parse them as [`LayerFileName`] then convert from that to a

View File

@@ -30,7 +30,6 @@ use crate::repository::{Key, Value, KEY_SIZE};
use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockReader, FileBlockReader};
use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
use crate::tenant::storage_layer::{
PersistentLayer, ValueReconstructResult, ValueReconstructState,
};
@@ -57,8 +56,8 @@ use utils::{
};
use super::{
DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, LayerFileName, LayerIter,
LayerKeyIter, PathOrConf, RemoteLayerDesc,
DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, LayerIter, LayerKeyIter,
PathOrConf, PersistentLayerDesc,
};
///
@@ -90,10 +89,10 @@ impl From<&DeltaLayer> for Summary {
magic: DELTA_FILE_MAGIC,
format_version: STORAGE_FORMAT_VERSION,
tenant_id: layer.tenant_id,
timeline_id: layer.timeline_id,
key_range: layer.key_range.clone(),
lsn_range: layer.lsn_range.clone(),
tenant_id: layer.desc.tenant_id,
timeline_id: layer.desc.timeline_id,
key_range: layer.desc.key_range.clone(),
lsn_range: layer.desc.lsn_range.clone(),
index_start_blk: 0,
index_root_blk: 0,
@@ -181,10 +180,7 @@ impl DeltaKey {
pub struct DeltaLayer {
path_or_conf: PathOrConf,
pub tenant_id: TenantId,
pub timeline_id: TimelineId,
pub key_range: Range<Key>,
pub lsn_range: Range<Lsn>,
pub desc: PersistentLayerDesc,
pub file_size: u64,
@@ -198,8 +194,8 @@ impl std::fmt::Debug for DeltaLayer {
use super::RangeDisplayDebug;
f.debug_struct("DeltaLayer")
.field("key_range", &RangeDisplayDebug(&self.key_range))
.field("lsn_range", &self.lsn_range)
.field("key_range", &RangeDisplayDebug(&self.desc.key_range))
.field("lsn_range", &self.desc.lsn_range)
.field("file_size", &self.file_size)
.field("inner", &self.inner)
.finish()
@@ -229,30 +225,16 @@ impl std::fmt::Debug for DeltaLayerInner {
}
impl Layer for DeltaLayer {
fn get_key_range(&self) -> Range<Key> {
self.key_range.clone()
}
fn get_lsn_range(&self) -> Range<Lsn> {
self.lsn_range.clone()
}
fn is_incremental(&self) -> bool {
true
}
fn short_id(&self) -> String {
self.filename().file_name()
}
/// debugging function to print out the contents of the layer
fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
println!(
"----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} ----",
self.tenant_id,
self.timeline_id,
self.key_range.start,
self.key_range.end,
self.lsn_range.start,
self.lsn_range.end
self.desc.tenant_id,
self.desc.timeline_id,
self.desc.key_range.start,
self.desc.key_range.end,
self.desc.lsn_range.start,
self.desc.lsn_range.end
);
if !verbose {
@@ -325,10 +307,10 @@ impl Layer for DeltaLayer {
reconstruct_state: &mut ValueReconstructState,
ctx: &RequestContext,
) -> anyhow::Result<ValueReconstructResult> {
ensure!(lsn_range.start >= self.lsn_range.start);
ensure!(lsn_range.start >= self.desc.lsn_range.start);
let mut need_image = true;
ensure!(self.key_range.contains(&key));
ensure!(self.desc.key_range.contains(&key));
{
// Open the file and lock the metadata in memory
@@ -403,19 +385,31 @@ impl Layer for DeltaLayer {
Ok(ValueReconstructResult::Complete)
}
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_key_range(&self) -> Range<Key> {
self.layer_desc().key_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_lsn_range(&self) -> Range<Lsn> {
self.layer_desc().lsn_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn is_incremental(&self) -> bool {
self.layer_desc().is_incremental
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn short_id(&self) -> String {
self.layer_desc().short_id()
}
}
impl PersistentLayer for DeltaLayer {
fn get_tenant_id(&self) -> TenantId {
self.tenant_id
}
fn get_timeline_id(&self) -> TimelineId {
self.timeline_id
}
fn filename(&self) -> LayerFileName {
self.layer_name().into()
fn layer_desc(&self) -> &PersistentLayerDesc {
&self.desc
}
fn local_path(&self) -> Option<PathBuf> {
@@ -603,10 +597,12 @@ impl DeltaLayer {
) -> DeltaLayer {
DeltaLayer {
path_or_conf: PathOrConf::Conf(conf),
timeline_id,
tenant_id,
key_range: filename.key_range.clone(),
lsn_range: filename.lsn_range.clone(),
desc: PersistentLayerDesc::new_delta(
tenant_id,
timeline_id,
filename.key_range.clone(),
filename.lsn_range.clone(),
),
file_size,
access_stats,
inner: RwLock::new(DeltaLayerInner {
@@ -633,10 +629,12 @@ impl DeltaLayer {
Ok(DeltaLayer {
path_or_conf: PathOrConf::Path(path.to_path_buf()),
timeline_id: summary.timeline_id,
tenant_id: summary.tenant_id,
key_range: summary.key_range,
lsn_range: summary.lsn_range,
desc: PersistentLayerDesc::new_delta(
summary.tenant_id,
summary.timeline_id,
summary.key_range,
summary.lsn_range,
),
file_size: metadata.len(),
access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
inner: RwLock::new(DeltaLayerInner {
@@ -649,32 +647,17 @@ impl DeltaLayer {
}
fn layer_name(&self) -> DeltaFileName {
DeltaFileName {
key_range: self.key_range.clone(),
lsn_range: self.lsn_range.clone(),
}
self.desc.delta_file_name()
}
/// Path to the layer file in pageserver workdir.
pub fn path(&self) -> PathBuf {
Self::path_for(
&self.path_or_conf,
self.timeline_id,
self.tenant_id,
self.desc.timeline_id,
self.desc.tenant_id,
&self.layer_name(),
)
}
/// Create layer descriptor for this image layer
pub fn layer_desc(&self) -> RemoteLayerDesc {
RemoteLayerDesc::new_delta(
self.tenant_id,
self.timeline_id,
&self.layer_name(),
&LayerFileMetadata::new(self.file_size()),
LayerAccessStats::empty_will_record_residence_event_later(),
)
}
}
/// A builder object for constructing a new delta layer.
@@ -815,10 +798,12 @@ impl DeltaLayerWriterInner {
// set inner.file here. The first read will have to re-open it.
let layer = DeltaLayer {
path_or_conf: PathOrConf::Conf(self.conf),
tenant_id: self.tenant_id,
timeline_id: self.timeline_id,
key_range: self.key_start..key_end,
lsn_range: self.lsn_range.clone(),
desc: PersistentLayerDesc::new_delta(
self.tenant_id,
self.timeline_id,
self.key_start..key_end,
self.lsn_range.clone(),
),
file_size: metadata.len(),
access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
inner: RwLock::new(DeltaLayerInner {

View File

@@ -9,6 +9,8 @@ use std::str::FromStr;
use utils::lsn::Lsn;
use super::PersistentLayerDesc;
// Note: Timeline::load_layer_map() relies on this sort order
#[derive(PartialEq, Eq, Clone, Hash)]
pub struct DeltaFileName {
@@ -153,7 +155,7 @@ impl Ord for ImageFileName {
impl ImageFileName {
pub fn lsn_as_range(&self) -> Range<Lsn> {
// Saves from having to copypaste this all over
self.lsn..(self.lsn + 1)
PersistentLayerDesc::image_layer_lsn_range(self.lsn)
}
}

View File

@@ -26,7 +26,6 @@ use crate::repository::{Key, KEY_SIZE};
use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
use crate::tenant::storage_layer::{
LayerAccessStats, PersistentLayer, ValueReconstructResult, ValueReconstructState,
};
@@ -53,8 +52,8 @@ use utils::{
lsn::Lsn,
};
use super::filename::{ImageFileName, LayerFileName};
use super::{Layer, LayerAccessStatsReset, LayerIter, PathOrConf, RemoteLayerDesc};
use super::filename::ImageFileName;
use super::{Layer, LayerAccessStatsReset, LayerIter, PathOrConf, PersistentLayerDesc};
///
/// Header stored in the beginning of the file
@@ -85,9 +84,9 @@ impl From<&ImageLayer> for Summary {
Self {
magic: IMAGE_FILE_MAGIC,
format_version: STORAGE_FORMAT_VERSION,
tenant_id: layer.tenant_id,
timeline_id: layer.timeline_id,
key_range: layer.key_range.clone(),
tenant_id: layer.desc.tenant_id,
timeline_id: layer.desc.timeline_id,
key_range: layer.desc.key_range.clone(),
lsn: layer.lsn,
index_start_blk: 0,
@@ -105,14 +104,13 @@ impl From<&ImageLayer> for Summary {
/// and it needs to be loaded before using it in queries.
pub struct ImageLayer {
path_or_conf: PathOrConf,
pub tenant_id: TenantId,
pub timeline_id: TimelineId,
pub key_range: Range<Key>,
pub file_size: u64,
// This entry contains an image of all pages as of this LSN
pub desc: PersistentLayerDesc,
// This entry contains an image of all pages as of this LSN, should be the same as desc.lsn
pub lsn: Lsn,
pub file_size: u64,
access_stats: LayerAccessStats,
inner: RwLock<ImageLayerInner>,
@@ -123,7 +121,7 @@ impl std::fmt::Debug for ImageLayer {
use super::RangeDisplayDebug;
f.debug_struct("ImageLayer")
.field("key_range", &RangeDisplayDebug(&self.key_range))
.field("key_range", &RangeDisplayDebug(&self.desc.key_range))
.field("file_size", &self.file_size)
.field("lsn", &self.lsn)
.field("inner", &self.inner)
@@ -154,27 +152,15 @@ impl std::fmt::Debug for ImageLayerInner {
}
impl Layer for ImageLayer {
fn get_key_range(&self) -> Range<Key> {
self.key_range.clone()
}
fn get_lsn_range(&self) -> Range<Lsn> {
// End-bound is exclusive
self.lsn..(self.lsn + 1)
}
fn is_incremental(&self) -> bool {
false
}
fn short_id(&self) -> String {
self.filename().file_name()
}
/// debugging function to print out the contents of the layer
fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
println!(
"----- image layer for ten {} tli {} key {}-{} at {} ----",
self.tenant_id, self.timeline_id, self.key_range.start, self.key_range.end, self.lsn
self.desc.tenant_id,
self.desc.timeline_id,
self.desc.key_range.start,
self.desc.key_range.end,
self.lsn
);
if !verbose {
@@ -204,7 +190,7 @@ impl Layer for ImageLayer {
reconstruct_state: &mut ValueReconstructState,
ctx: &RequestContext,
) -> anyhow::Result<ValueReconstructResult> {
assert!(self.key_range.contains(&key));
assert!(self.desc.key_range.contains(&key));
assert!(lsn_range.start >= self.lsn);
assert!(lsn_range.end >= self.lsn);
@@ -231,24 +217,37 @@ impl Layer for ImageLayer {
Ok(ValueReconstructResult::Missing)
}
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_key_range(&self) -> Range<Key> {
self.layer_desc().key_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_lsn_range(&self) -> Range<Lsn> {
self.layer_desc().lsn_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn is_incremental(&self) -> bool {
self.layer_desc().is_incremental
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn short_id(&self) -> String {
self.layer_desc().short_id()
}
}
impl PersistentLayer for ImageLayer {
fn filename(&self) -> LayerFileName {
self.layer_name().into()
fn layer_desc(&self) -> &PersistentLayerDesc {
&self.desc
}
fn local_path(&self) -> Option<PathBuf> {
Some(self.path())
}
fn get_tenant_id(&self) -> TenantId {
self.tenant_id
}
fn get_timeline_id(&self) -> TimelineId {
self.timeline_id
}
fn iter(&self, _ctx: &RequestContext) -> Result<LayerIter<'_>> {
unimplemented!();
}
@@ -406,9 +405,13 @@ impl ImageLayer {
) -> ImageLayer {
ImageLayer {
path_or_conf: PathOrConf::Conf(conf),
timeline_id,
tenant_id,
key_range: filename.key_range.clone(),
desc: PersistentLayerDesc::new_img(
tenant_id,
timeline_id,
filename.key_range.clone(),
filename.lsn,
false,
), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
lsn: filename.lsn,
file_size,
access_stats,
@@ -434,9 +437,13 @@ impl ImageLayer {
.context("get file metadata to determine size")?;
Ok(ImageLayer {
path_or_conf: PathOrConf::Path(path.to_path_buf()),
timeline_id: summary.timeline_id,
tenant_id: summary.tenant_id,
key_range: summary.key_range,
desc: PersistentLayerDesc::new_img(
summary.tenant_id,
summary.timeline_id,
summary.key_range,
summary.lsn,
false,
), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
lsn: summary.lsn,
file_size: metadata.len(),
access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
@@ -450,32 +457,18 @@ impl ImageLayer {
}
fn layer_name(&self) -> ImageFileName {
ImageFileName {
key_range: self.key_range.clone(),
lsn: self.lsn,
}
self.desc.image_file_name()
}
/// Path to the layer file in pageserver workdir.
pub fn path(&self) -> PathBuf {
Self::path_for(
&self.path_or_conf,
self.timeline_id,
self.tenant_id,
self.desc.timeline_id,
self.desc.tenant_id,
&self.layer_name(),
)
}
/// Create layer descriptor for this image layer
pub fn layer_desc(&self) -> RemoteLayerDesc {
RemoteLayerDesc::new_img(
self.tenant_id,
self.timeline_id,
&self.layer_name(),
&LayerFileMetadata::new(self.file_size()),
LayerAccessStats::empty_will_record_residence_event_later(),
)
}
}
/// A builder object for constructing a new image layer.
@@ -496,6 +489,7 @@ struct ImageLayerWriterInner {
tenant_id: TenantId,
key_range: Range<Key>,
lsn: Lsn,
is_incremental: bool,
blob_writer: WriteBlobWriter<VirtualFile>,
tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
@@ -511,6 +505,7 @@ impl ImageLayerWriterInner {
tenant_id: TenantId,
key_range: &Range<Key>,
lsn: Lsn,
is_incremental: bool,
) -> anyhow::Result<Self> {
// Create the file initially with a temporary filename.
// We'll atomically rename it to the final name when we're done.
@@ -545,6 +540,7 @@ impl ImageLayerWriterInner {
lsn,
tree: tree_builder,
blob_writer,
is_incremental,
};
Ok(writer)
@@ -582,6 +578,14 @@ impl ImageLayerWriterInner {
file.write_all(buf.as_ref())?;
}
let desc = PersistentLayerDesc::new_img(
self.tenant_id,
self.timeline_id,
self.key_range.clone(),
self.lsn,
self.is_incremental, // for now, image layer ALWAYS covers the full range
);
// Fill in the summary on blk 0
let summary = Summary {
magic: IMAGE_FILE_MAGIC,
@@ -605,9 +609,7 @@ impl ImageLayerWriterInner {
// set inner.file here. The first read will have to re-open it.
let layer = ImageLayer {
path_or_conf: PathOrConf::Conf(self.conf),
timeline_id: self.timeline_id,
tenant_id: self.tenant_id,
key_range: self.key_range.clone(),
desc,
lsn: self.lsn,
file_size: metadata.len(),
access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
@@ -679,6 +681,7 @@ impl ImageLayerWriter {
tenant_id: TenantId,
key_range: &Range<Key>,
lsn: Lsn,
is_incremental: bool,
) -> anyhow::Result<ImageLayerWriter> {
Ok(Self {
inner: Some(ImageLayerWriterInner::new(
@@ -687,6 +690,7 @@ impl ImageLayerWriter {
tenant_id,
key_range,
lsn,
is_incremental,
)?),
})
}

View File

@@ -0,0 +1,109 @@
use std::ops::Range;
use utils::{
id::{TenantId, TimelineId},
lsn::Lsn,
};
use crate::repository::Key;
use super::{DeltaFileName, ImageFileName, LayerFileName};
/// A unique identifier of a persistent layer. This is different from `LayerDescriptor`, which is only used in the
/// benchmarks. This struct contains all necessary information to find the image / delta layer. It also provides
/// a unified way to generate layer information like file name.
#[derive(Debug, PartialEq, Eq, Clone)]
pub struct PersistentLayerDesc {
pub tenant_id: TenantId,
pub timeline_id: TimelineId,
pub key_range: Range<Key>,
/// For image layer, this is `[lsn, lsn+1)`.
pub lsn_range: Range<Lsn>,
/// Whether this is a delta layer.
pub is_delta: bool,
/// Whether this layer only contains page images for part of the keys in the range. In the current implementation, this should
/// always be equal to `is_delta`. If we land the partial image layer PR someday, image layer could also be
/// incremental.
pub is_incremental: bool,
}
impl PersistentLayerDesc {
pub fn short_id(&self) -> String {
self.filename().file_name()
}
pub fn new_img(
tenant_id: TenantId,
timeline_id: TimelineId,
key_range: Range<Key>,
lsn: Lsn,
is_incremental: bool,
) -> Self {
Self {
tenant_id,
timeline_id,
key_range,
lsn_range: Self::image_layer_lsn_range(lsn),
is_delta: false,
is_incremental,
}
}
pub fn new_delta(
tenant_id: TenantId,
timeline_id: TimelineId,
key_range: Range<Key>,
lsn_range: Range<Lsn>,
) -> Self {
Self {
tenant_id,
timeline_id,
key_range,
lsn_range,
is_delta: true,
is_incremental: true,
}
}
/// Get the LSN that the image layer covers.
pub fn image_layer_lsn(&self) -> Lsn {
assert!(!self.is_delta);
assert!(self.lsn_range.start + 1 == self.lsn_range.end);
self.lsn_range.start
}
/// Get the LSN range corresponding to a single image layer LSN.
pub fn image_layer_lsn_range(lsn: Lsn) -> Range<Lsn> {
lsn..(lsn + 1)
}
/// Get a delta file name for this layer.
///
/// Panic: if this is not a delta layer.
pub fn delta_file_name(&self) -> DeltaFileName {
assert!(self.is_delta);
DeltaFileName {
key_range: self.key_range.clone(),
lsn_range: self.lsn_range.clone(),
}
}
/// Get a delta file name for this layer.
///
/// Panic: if this is not an image layer, or the lsn range is invalid
pub fn image_file_name(&self) -> ImageFileName {
assert!(!self.is_delta);
assert!(self.lsn_range.start + 1 == self.lsn_range.end);
ImageFileName {
key_range: self.key_range.clone(),
lsn: self.lsn_range.start,
}
}
pub fn filename(&self) -> LayerFileName {
if self.is_delta {
self.delta_file_name().into()
} else {
self.image_file_name().into()
}
}
}

View File

@@ -1,4 +1,4 @@
//! A RemoteLayerDesc is an in-memory placeholder for a layer file that exists
//! A RemoteLayer is an in-memory placeholder for a layer file that exists
//! in remote storage.
//!
use crate::config::PageServerConf;
@@ -18,35 +18,25 @@ use utils::{
lsn::Lsn,
};
use super::filename::{DeltaFileName, ImageFileName, LayerFileName};
use super::image_layer::ImageLayer;
use super::filename::{DeltaFileName, ImageFileName};
use super::{
DeltaLayer, LayerAccessStats, LayerAccessStatsReset, LayerIter, LayerKeyIter,
LayerResidenceStatus, PersistentLayer,
DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset, LayerIter, LayerKeyIter,
LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
};
/// RemoteLayerDesc is a not yet downloaded [`ImageLayer`] or
/// RemoteLayer is a not yet downloaded [`ImageLayer`] or
/// [`crate::storage_layer::DeltaLayer`].
///
/// RemoteLayerDesc might be downloaded on-demand during operations which are
/// RemoteLayer might be downloaded on-demand during operations which are
/// allowed download remote layers and during which, it gets replaced with a
/// concrete `DeltaLayer` or `ImageLayer`.
///
/// See: [`crate::context::RequestContext`] for authorization to download
pub struct RemoteLayerDesc {
pub(crate) tenantid: TenantId,
pub(crate) timelineid: TimelineId,
pub(crate) key_range: Range<Key>,
pub(crate) lsn_range: Range<Lsn>,
pub file_name: LayerFileName,
pub struct RemoteLayer {
pub desc: PersistentLayerDesc,
pub layer_metadata: LayerFileMetadata,
is_delta: bool,
is_incremental: bool,
access_stats: LayerAccessStats,
pub(crate) ongoing_download: Arc<tokio::sync::Semaphore>,
@@ -54,7 +44,7 @@ pub struct RemoteLayerDesc {
/// Has `LayerMap::replace` failed for this (true) or not (false).
///
/// Used together with [`ongoing_download`] semaphore in `Timeline::download_remote_layer`.
/// The field is used to mark a RemoteLayerDesc permanently (until restart or ignore+load)
/// The field is used to mark a RemoteLayer permanently (until restart or ignore+load)
/// unprocessable, because a LayerMap::replace failed.
///
/// It is very unlikely to accumulate these in the Timeline's LayerMap, but having this avoids
@@ -63,25 +53,17 @@ pub struct RemoteLayerDesc {
pub(crate) download_replacement_failure: std::sync::atomic::AtomicBool,
}
impl std::fmt::Debug for RemoteLayerDesc {
impl std::fmt::Debug for RemoteLayer {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
f.debug_struct("RemoteLayerDesc")
.field("file_name", &self.file_name)
f.debug_struct("RemoteLayer")
.field("file_name", &self.desc.filename())
.field("layer_metadata", &self.layer_metadata)
.field("is_incremental", &self.is_incremental)
.field("is_incremental", &self.desc.is_incremental)
.finish()
}
}
impl Layer for RemoteLayerDesc {
fn get_key_range(&self) -> Range<Key> {
self.key_range.clone()
}
fn get_lsn_range(&self) -> Range<Lsn> {
self.lsn_range.clone()
}
impl Layer for RemoteLayer {
fn get_value_reconstruct_data(
&self,
_key: Key,
@@ -95,53 +77,45 @@ impl Layer for RemoteLayerDesc {
);
}
fn is_incremental(&self) -> bool {
self.is_incremental
}
/// debugging function to print out the contents of the layer
fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
println!(
"----- remote layer for ten {} tli {} keys {}-{} lsn {}-{} ----",
self.tenantid,
self.timelineid,
self.key_range.start,
self.key_range.end,
self.lsn_range.start,
self.lsn_range.end
self.desc.tenant_id,
self.desc.timeline_id,
self.desc.key_range.start,
self.desc.key_range.end,
self.desc.lsn_range.start,
self.desc.lsn_range.end
);
Ok(())
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_key_range(&self) -> Range<Key> {
self.layer_desc().key_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn get_lsn_range(&self) -> Range<Lsn> {
self.layer_desc().lsn_range.clone()
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn is_incremental(&self) -> bool {
self.layer_desc().is_incremental
}
/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
fn short_id(&self) -> String {
self.filename().file_name()
self.layer_desc().short_id()
}
}
impl PersistentLayer for RemoteLayerDesc {
fn get_tenant_id(&self) -> TenantId {
self.tenantid
}
fn get_timeline_id(&self) -> TimelineId {
self.timelineid
}
fn filename(&self) -> LayerFileName {
if self.is_delta {
DeltaFileName {
key_range: self.key_range.clone(),
lsn_range: self.lsn_range.clone(),
}
.into()
} else {
ImageFileName {
key_range: self.key_range.clone(),
lsn: self.lsn_range.start,
}
.into()
}
impl PersistentLayer for RemoteLayer {
fn layer_desc(&self) -> &PersistentLayerDesc {
&self.desc
}
fn local_path(&self) -> Option<PathBuf> {
@@ -160,6 +134,14 @@ impl PersistentLayer for RemoteLayerDesc {
bail!("remote layer has no layer file");
}
fn downcast_remote_layer<'a>(self: Arc<Self>) -> Option<std::sync::Arc<RemoteLayer>> {
Some(self)
}
fn is_remote_layer(&self) -> bool {
true
}
fn file_size(&self) -> u64 {
self.layer_metadata.file_size()
}
@@ -168,7 +150,7 @@ impl PersistentLayer for RemoteLayerDesc {
let layer_file_name = self.filename().file_name();
let lsn_range = self.get_lsn_range();
if self.is_delta {
if self.desc.is_delta {
HistoricLayerInfo::Delta {
layer_file_name,
layer_file_size: self.layer_metadata.file_size(),
@@ -193,22 +175,22 @@ impl PersistentLayer for RemoteLayerDesc {
}
}
impl RemoteLayerDesc {
impl RemoteLayer {
pub fn new_img(
tenantid: TenantId,
timelineid: TimelineId,
fname: &ImageFileName,
layer_metadata: &LayerFileMetadata,
access_stats: LayerAccessStats,
) -> RemoteLayerDesc {
RemoteLayerDesc {
tenantid,
timelineid,
key_range: fname.key_range.clone(),
lsn_range: fname.lsn_as_range(),
is_delta: false,
is_incremental: false,
file_name: fname.to_owned().into(),
) -> RemoteLayer {
RemoteLayer {
desc: PersistentLayerDesc::new_img(
tenantid,
timelineid,
fname.key_range.clone(),
fname.lsn,
false,
),
layer_metadata: layer_metadata.clone(),
ongoing_download: Arc::new(tokio::sync::Semaphore::new(1)),
download_replacement_failure: std::sync::atomic::AtomicBool::default(),
@@ -222,15 +204,14 @@ impl RemoteLayerDesc {
fname: &DeltaFileName,
layer_metadata: &LayerFileMetadata,
access_stats: LayerAccessStats,
) -> RemoteLayerDesc {
RemoteLayerDesc {
tenantid,
timelineid,
key_range: fname.key_range.clone(),
lsn_range: fname.lsn_range.clone(),
is_delta: true,
is_incremental: true,
file_name: fname.to_owned().into(),
) -> RemoteLayer {
RemoteLayer {
desc: PersistentLayerDesc::new_delta(
tenantid,
timelineid,
fname.key_range.clone(),
fname.lsn_range.clone(),
),
layer_metadata: layer_metadata.clone(),
ongoing_download: Arc::new(tokio::sync::Semaphore::new(1)),
download_replacement_failure: std::sync::atomic::AtomicBool::default(),
@@ -248,15 +229,12 @@ impl RemoteLayerDesc {
where
L: ?Sized + Layer,
{
if self.is_delta {
let fname = DeltaFileName {
key_range: self.key_range.clone(),
lsn_range: self.lsn_range.clone(),
};
if self.desc.is_delta {
let fname = self.desc.delta_file_name();
Arc::new(DeltaLayer::new(
conf,
self.timelineid,
self.tenantid,
self.desc.timeline_id,
self.desc.tenant_id,
&fname,
file_size,
self.access_stats.clone_for_residence_change(
@@ -265,14 +243,11 @@ impl RemoteLayerDesc {
),
))
} else {
let fname = ImageFileName {
key_range: self.key_range.clone(),
lsn: self.lsn_range.start,
};
let fname = self.desc.image_file_name();
Arc::new(ImageLayer::new(
conf,
self.timelineid,
self.tenantid,
self.desc.timeline_id,
self.desc.tenant_id,
&fname,
file_size,
self.access_stats.clone_for_residence_change(

View File

@@ -35,7 +35,7 @@ use crate::context::{DownloadBehavior, RequestContext};
use crate::tenant::remote_timeline_client::{self, index::LayerFileMetadata};
use crate::tenant::storage_layer::{
DeltaFileName, DeltaLayerWriter, ImageFileName, ImageLayerWriter, InMemoryLayer,
LayerAccessStats, LayerFileName, RemoteLayerDesc,
LayerAccessStats, LayerFileName, RemoteLayer,
};
use crate::tenant::{
ephemeral_file::is_ephemeral_file,
@@ -77,7 +77,6 @@ use self::eviction_task::EvictionTaskTimelineState;
use self::walreceiver::{WalReceiver, WalReceiverConf};
use super::config::TenantConf;
use super::layer_cache::LayerCache;
use super::layer_map::BatchedUpdates;
use super::remote_timeline_client::index::IndexPart;
use super::remote_timeline_client::RemoteTimelineClient;
@@ -120,7 +119,7 @@ pub struct Timeline {
pub pg_version: u32,
pub(super) layers: RwLock<LayerMap<RemoteLayerDesc>>,
pub(super) layers: RwLock<LayerMap<dyn PersistentLayer>>,
/// Set of key ranges which should be covered by image layers to
/// allow GC to remove old layers. This set is created by GC and its cutoff LSN is also stored.
@@ -242,8 +241,6 @@ pub struct Timeline {
pub delete_lock: tokio::sync::Mutex<bool>,
eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
layer_cache: Arc<LayerCache>,
}
/// Internal structure to hold all data needed for logical size calculation.
@@ -528,7 +525,12 @@ impl Timeline {
Some((cached_lsn, cached_img)) => {
match cached_lsn.cmp(&lsn) {
Ordering::Less => {} // there might be WAL between cached_lsn and lsn, we need to check
Ordering::Equal => return Ok(cached_img), // exact LSN match, return the image
Ordering::Equal => {
self.metrics
.materialized_page_cache_hit_upon_request_counter
.inc();
return Ok(cached_img); // exact LSN match, return the image
}
Ordering::Greater => {
unreachable!("the returned lsn should never be after the requested lsn")
}
@@ -543,8 +545,10 @@ impl Timeline {
img: cached_page_img,
};
let timer = self.metrics.get_reconstruct_data_time_histo.start_timer();
self.get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
.await?;
timer.stop_and_record();
self.metrics
.reconstruct_time_histo
@@ -1010,22 +1014,20 @@ impl Timeline {
#[instrument(skip_all, fields(tenant = %self.tenant_id, timeline = %self.timeline_id))]
pub async fn download_layer(&self, layer_file_name: &str) -> anyhow::Result<Option<bool>> {
let Some(remote_layer_desc) = self.find_layer_desc(layer_file_name) else { return Ok(None) };
if self.layer_cache.contains(&remote_layer_desc.filename()) {
return Ok(Some(false));
}
let Some(layer) = self.find_layer(layer_file_name) else { return Ok(None) };
let Some(remote_layer) = layer.downcast_remote_layer() else { return Ok(Some(false)) };
if self.remote_client.is_none() {
return Ok(Some(false));
}
self.download_remote_layer(remote_layer_desc).await?;
self.download_remote_layer(remote_layer).await?;
Ok(Some(true))
}
/// Like [`evict_layer_batch`], but for just one layer.
/// Additional case `Ok(None)` covers the case where the layer could not be found by its `layer_file_name`.
pub async fn evict_layer(&self, layer_file_name: &str) -> anyhow::Result<Option<bool>> {
let Some(local_layer) = self.find_layer_desc(layer_file_name) else { return Ok(None) };
let Some(local_layer) = self.find_layer(layer_file_name) else { return Ok(None) };
let remote_client = self
.remote_client
.as_ref()
@@ -1052,7 +1054,7 @@ impl Timeline {
pub async fn evict_layers(
&self,
_: &GenericRemoteStorage,
layers_to_evict: &[Arc<RemoteLayerDesc>],
layers_to_evict: &[Arc<dyn PersistentLayer>],
cancel: CancellationToken,
) -> anyhow::Result<Vec<Option<anyhow::Result<bool>>>> {
let remote_client = self.remote_client.clone().expect(
@@ -1087,7 +1089,7 @@ impl Timeline {
async fn evict_layer_batch(
&self,
remote_client: &Arc<RemoteTimelineClient>,
layers_to_evict: &[Arc<RemoteLayerDesc>],
layers_to_evict: &[Arc<dyn PersistentLayer>],
cancel: CancellationToken,
) -> anyhow::Result<Vec<Option<anyhow::Result<bool>>>> {
// ensure that the layers have finished uploading
@@ -1136,12 +1138,12 @@ impl Timeline {
fn evict_layer_batch_impl(
&self,
_layer_removal_cs: &tokio::sync::MutexGuard<'_, ()>,
local_layer: &Arc<RemoteLayerDesc>,
batch_updates: &mut BatchedUpdates<'_, RemoteLayerDesc>,
local_layer: &Arc<dyn PersistentLayer>,
batch_updates: &mut BatchedUpdates<'_, dyn PersistentLayer>,
) -> anyhow::Result<bool> {
use super::layer_map::Replacement;
if !self.layer_cache.contains(&local_layer.filename()) {
if local_layer.is_remote_layer() {
// TODO(issue #3851): consider returning an err here instead of false,
// which is the same out the match later
return Ok(false);
@@ -1168,7 +1170,7 @@ impl Timeline {
let layer_metadata = LayerFileMetadata::new(layer_file_size);
let new_remote_layer = Arc::new(match local_layer.filename() {
LayerFileName::Image(image_name) => RemoteLayerDesc::new_img(
LayerFileName::Image(image_name) => RemoteLayer::new_img(
self.tenant_id,
self.timeline_id,
&image_name,
@@ -1177,7 +1179,7 @@ impl Timeline {
.access_stats()
.clone_for_residence_change(batch_updates, LayerResidenceStatus::Evicted),
),
LayerFileName::Delta(delta_name) => RemoteLayerDesc::new_delta(
LayerFileName::Delta(delta_name) => RemoteLayer::new_delta(
self.tenant_id,
self.timeline_id,
&delta_name,
@@ -1188,7 +1190,6 @@ impl Timeline {
),
});
/*
let replaced = match batch_updates.replace_historic(local_layer, new_remote_layer)? {
Replacement::Replaced { .. } => {
if let Err(e) = local_layer.delete_resident_layer_file() {
@@ -1239,10 +1240,8 @@ impl Timeline {
false
}
};
*/
// Ok(replaced)
Ok(true)
Ok(replaced)
}
}
@@ -1299,6 +1298,13 @@ impl Timeline {
.unwrap_or(default_tenant_conf.evictions_low_residence_duration_metric_threshold)
}
fn get_gc_feedback(&self) -> bool {
let tenant_conf = self.tenant_conf.read().unwrap();
tenant_conf
.gc_feedback
.unwrap_or(self.conf.default_tenant_conf.gc_feedback)
}
pub(super) fn tenant_conf_updated(&self) {
// NB: Most tenant conf options are read by background loops, so,
// changes will automatically be picked up.
@@ -1427,8 +1433,6 @@ impl Timeline {
EvictionTaskTimelineState::default(),
),
delete_lock: tokio::sync::Mutex::new(false),
layer_cache: Arc::new(LayerCache::new()),
};
result.repartition_threshold = result.get_checkpoint_distance() / 10;
result
@@ -1575,12 +1579,9 @@ impl Timeline {
LayerAccessStats::for_loading_layer(&updates, LayerResidenceStatus::Resident),
);
let remote_desc = layer.layer_desc();
trace!("found layer {}", layer.path().display());
total_physical_size += file_size;
self.layer_cache.insert(layer.filename(), Arc::new(layer));
updates.insert_historic(Arc::new(remote_desc));
updates.insert_historic(Arc::new(layer));
num_layers += 1;
} else if let Some(deltafilename) = DeltaFileName::parse_str(&fname) {
// Create a DeltaLayer struct for each delta file.
@@ -1612,9 +1613,7 @@ impl Timeline {
trace!("found layer {}", layer.path().display());
total_physical_size += file_size;
let remote_desc = layer.layer_desc();
self.layer_cache.insert(layer.filename(), Arc::new(layer));
updates.insert_historic(Arc::new(remote_desc));
updates.insert_historic(Arc::new(layer));
num_layers += 1;
} else if fname == METADATA_FILE_NAME || fname.ends_with(".old") {
// ignore these
@@ -1659,9 +1658,9 @@ impl Timeline {
async fn create_remote_layers(
&self,
index_part: &IndexPart,
local_layers: HashMap<LayerFileName, Arc<RemoteLayerDesc>>,
local_layers: HashMap<LayerFileName, Arc<dyn PersistentLayer>>,
up_to_date_disk_consistent_lsn: Lsn,
) -> anyhow::Result<HashMap<LayerFileName, Arc<RemoteLayerDesc>>> {
) -> anyhow::Result<HashMap<LayerFileName, Arc<dyn PersistentLayer>>> {
// Are we missing some files that are present in remote storage?
// Create RemoteLayer instances for them.
let mut local_only_layers = local_layers;
@@ -1740,7 +1739,7 @@ impl Timeline {
continue;
}
let remote_layer = RemoteLayerDesc::new_img(
let remote_layer = RemoteLayer::new_img(
self.tenant_id,
self.timeline_id,
imgfilename,
@@ -1768,7 +1767,7 @@ impl Timeline {
);
continue;
}
let remote_layer = RemoteLayerDesc::new_delta(
let remote_layer = RemoteLayer::new_delta(
self.tenant_id,
self.timeline_id,
deltafilename,
@@ -2174,7 +2173,7 @@ impl Timeline {
}
}
fn find_layer_desc(&self, layer_file_name: &str) -> Option<Arc<RemoteLayerDesc>> {
fn find_layer(&self, layer_file_name: &str) -> Option<Arc<dyn PersistentLayer>> {
for historic_layer in self.layers.read().unwrap().iter_historic_layers() {
let historic_layer_name = historic_layer.filename().file_name();
if layer_file_name == historic_layer_name {
@@ -2191,10 +2190,10 @@ impl Timeline {
&self,
// we cannot remove layers otherwise, since gc and compaction will race
_layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
layer: Arc<RemoteLayerDesc>,
updates: &mut BatchedUpdates<'_, RemoteLayerDesc>,
layer: Arc<dyn PersistentLayer>,
updates: &mut BatchedUpdates<'_, dyn PersistentLayer>,
) -> anyhow::Result<()> {
if self.layer_cache.contains(&layer.filename()) {
if !layer.is_remote_layer() {
layer.delete_resident_layer_file()?;
let layer_file_size = layer.file_size();
self.metrics
@@ -2269,6 +2268,9 @@ impl Timeline {
let mut timeline_owned;
let mut timeline = self;
let mut read_count =
scopeguard::guard(0, |cnt| self.metrics.read_num_fs_layers.observe(cnt as f64));
// For debugging purposes, collect the path of layers that we traversed
// through. It's included in the error message if we fail to find the key.
let mut traversal_path = Vec::<TraversalPathItem>::new();
@@ -2403,6 +2405,7 @@ impl Timeline {
Err(e) => return Err(PageReconstructError::from(e)),
};
cont_lsn = lsn_floor;
// metrics: open_layer does not count as fs access, so we are not updating `read_count`
traversal_path.push((
result,
cont_lsn,
@@ -2429,6 +2432,7 @@ impl Timeline {
Err(e) => return Err(PageReconstructError::from(e)),
};
cont_lsn = lsn_floor;
// metrics: open_layer does not count as fs access, so we are not updating `read_count`
traversal_path.push((
result,
cont_lsn,
@@ -2443,7 +2447,13 @@ impl Timeline {
if let Some(SearchResult { lsn_floor, layer }) = layers.search(key, cont_lsn) {
// If it's a remote layer, download it and retry.
if let Some(layer) = self.layer_cache.get(&layer.filename()) {
if let Some(remote_layer) =
super::storage_layer::downcast_remote_layer(&layer)
{
// TODO: push a breadcrumb to 'traversal_path' to record the fact that
// we downloaded / would need to download this layer.
remote_layer // download happens outside the scope of `layers` guard object
} else {
// Get all the data needed to reconstruct the page version from this layer.
// But if we have an older cached page image, no need to go past that.
let lsn_floor = max(cached_lsn + 1, lsn_floor);
@@ -2457,6 +2467,7 @@ impl Timeline {
Err(e) => return Err(PageReconstructError::from(e)),
};
cont_lsn = lsn_floor;
*read_count += 1;
traversal_path.push((
result,
cont_lsn,
@@ -2466,10 +2477,6 @@ impl Timeline {
}),
));
continue 'outer;
} else {
// TODO: push a breadcrumb to 'traversal_path' to record the fact that
// we downloaded / would need to download this layer.
layer // download happens outside the scope of `layers` guard object
}
} else if timeline.ancestor_timeline.is_some() {
// Nothing on this timeline. Traverse to parent
@@ -2526,7 +2533,7 @@ impl Timeline {
(DownloadBehavior::Error, false) => {
return Err(PageReconstructError::NeedsDownload(
TenantTimelineId::new(self.tenant_id, self.timeline_id),
remote_layer.file_name.clone(),
remote_layer.filename(),
))
}
}
@@ -2909,8 +2916,7 @@ impl Timeline {
LayerResidenceStatus::Resident,
LayerResidenceEventReason::LayerCreate,
);
batch_updates.insert_historic(Arc::new(l.layer_desc()));
self.layer_cache.insert(l.filename(), l);
batch_updates.insert_historic(l);
batch_updates.flush();
// update the timeline's physical size
@@ -3060,6 +3066,7 @@ impl Timeline {
self.tenant_id,
&img_range,
lsn,
false, // image layer always covers the full range
)?;
fail_point!("image-layer-writer-fail-before-finish", |_| {
@@ -3138,6 +3145,7 @@ impl Timeline {
let mut layers = self.layers.write().unwrap();
let mut updates = layers.batch_update();
let timeline_path = self.conf.timeline_path(&self.timeline_id, &self.tenant_id);
for l in image_layers {
let path = l.filename();
let metadata = timeline_path
@@ -3156,9 +3164,7 @@ impl Timeline {
LayerResidenceStatus::Resident,
LayerResidenceEventReason::LayerCreate,
);
updates.insert_historic(Arc::new(l.layer_desc()));
let x: Arc<dyn PersistentLayer> = l;
self.layer_cache.insert(x.filename(), x)
updates.insert_historic(l);
}
updates.flush();
drop(layers);
@@ -3171,7 +3177,7 @@ impl Timeline {
#[derive(Default)]
struct CompactLevel0Phase1Result {
new_layers: Vec<DeltaLayer>,
deltas_to_compact: Vec<Arc<RemoteLayerDesc>>,
deltas_to_compact: Vec<Arc<dyn PersistentLayer>>,
}
/// Top-level failure to compact.
@@ -3181,7 +3187,7 @@ enum CompactionError {
///
/// This should not happen repeatedly, but will be retried once by top-level
/// `Timeline::compact`.
DownloadRequired(Vec<Arc<RemoteLayerDesc>>),
DownloadRequired(Vec<Arc<RemoteLayer>>),
/// Compaction cannot be done right now; page reconstruction and so on.
Other(anyhow::Error),
}
@@ -3253,9 +3259,13 @@ impl Timeline {
let remotes = deltas_to_compact
.iter()
.filter(|l| !self.layer_cache.contains(&l.filename()))
.filter(|l| l.is_remote_layer())
.inspect(|l| info!("compact requires download of {}", l.filename().file_name()))
.cloned()
.map(|l| {
l.clone()
.downcast_remote_layer()
.expect("just checked it is remote layer")
})
.collect::<Vec<_>>();
if !remotes.is_empty() {
@@ -3595,15 +3605,13 @@ impl Timeline {
.add(metadata.len());
new_layer_paths.insert(new_delta_path, LayerFileMetadata::new(metadata.len()));
let remote_desc = l.layer_desc();
let x: Arc<dyn PersistentLayer + 'static> = Arc::new(l);
x.access_stats().record_residence_event(
&updates,
LayerResidenceStatus::Resident,
LayerResidenceEventReason::LayerCreate,
);
updates.insert_historic(Arc::new(remote_desc));
self.layer_cache.insert(x.filename(), x)
updates.insert_historic(x);
}
// Now that we have reshuffled the data to set of new delta layers, we can
@@ -3910,7 +3918,7 @@ impl Timeline {
// delta layers. Image layers can form "stairs" preventing old image from been deleted.
// But image layers are in any case less sparse than delta layers. Also we need some
// protection from replacing recent image layers with new one after each GC iteration.
if l.is_incremental() && !LayerMap::is_l0(&*l) {
if self.get_gc_feedback() && l.is_incremental() && !LayerMap::is_l0(&*l) {
wanted_image_layers.add_range(l.get_key_range());
}
result.layers_not_updated += 1;
@@ -4076,7 +4084,7 @@ impl Timeline {
#[instrument(skip_all, fields(layer=%remote_layer.short_id()))]
pub async fn download_remote_layer(
&self,
remote_layer: Arc<RemoteLayerDesc>,
remote_layer: Arc<RemoteLayer>,
) -> anyhow::Result<()> {
debug_assert_current_span_has_tenant_and_timeline_id();
@@ -4119,7 +4127,7 @@ impl Timeline {
// Does retries + exponential back-off internally.
// When this fails, don't layer further retry attempts here.
let result = remote_client
.download_layer_file(&remote_layer.file_name, &remote_layer.layer_metadata)
.download_layer_file(&remote_layer.filename(), &remote_layer.layer_metadata)
.await;
if let Ok(size) = &result {
@@ -4133,12 +4141,10 @@ impl Timeline {
// Delta- or ImageLayer in the layer map.
let mut layers = self_clone.layers.write().unwrap();
let mut updates = layers.batch_update();
let new_layer =
remote_layer.create_downloaded_layer(&updates, self_clone.conf, *size);
let new_layer = remote_layer.create_downloaded_layer(&updates, self_clone.conf, *size);
{
use crate::tenant::layer_map::Replacement;
let l: Arc<dyn PersistentLayer> = remote_layer.clone();
/*
let failure = match updates.replace_historic(&l, new_layer) {
Ok(Replacement::Replaced { .. }) => false,
Ok(Replacement::NotFound) => {
@@ -4193,9 +4199,8 @@ impl Timeline {
remote_layer
.download_replacement_failure
.store(true, Relaxed);
} */
}
}
updates.flush();
drop(layers);
@@ -4208,10 +4213,7 @@ impl Timeline {
remote_layer.ongoing_download.close();
} else {
// Keep semaphore open. We'll drop the permit at the end of the function.
error!(
"layer file download failed: {:?}",
result.as_ref().unwrap_err()
);
error!("layer file download failed: {:?}", result.as_ref().unwrap_err());
}
// Don't treat it as an error if the task that triggered the download
@@ -4225,8 +4227,7 @@ impl Timeline {
drop(permit);
Ok(())
}
.in_current_span(),
}.in_current_span(),
);
receiver.await.context("download task cancelled")?
@@ -4299,7 +4300,7 @@ impl Timeline {
let layers = self.layers.read().unwrap();
layers
.iter_historic_layers()
.filter(|l| !self.layer_cache.contains(&l.filename()))
.filter_map(|l| l.downcast_remote_layer())
.map(|l| self.download_remote_layer(l))
.for_each(|dl| downloads.push(dl))
}
@@ -4374,7 +4375,7 @@ pub struct DiskUsageEvictionInfo {
}
pub struct LocalLayerInfoForDiskUsageEviction {
pub layer: Arc<RemoteLayerDesc>,
pub layer: Arc<dyn PersistentLayer>,
pub last_activity_ts: SystemTime,
}
@@ -4408,7 +4409,7 @@ impl Timeline {
let file_size = l.file_size();
max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
if !self.layer_cache.contains(&l.filename()) {
if l.is_remote_layer() {
continue;
}

View File

@@ -29,7 +29,7 @@ use crate::{
task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
tenant::{
config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
storage_layer::{PersistentLayer, RemoteLayerDesc},
storage_layer::PersistentLayer,
LogicalSizeCalculationCause, Tenant,
},
};
@@ -184,11 +184,11 @@ impl Timeline {
// NB: all the checks can be invalidated as soon as we release the layer map lock.
// We don't want to hold the layer map lock during eviction.
// So, we just need to deal with this.
let candidates: Vec<Arc<RemoteLayerDesc>> = {
let candidates: Vec<Arc<dyn PersistentLayer>> = {
let layers = self.layers.read().unwrap();
let mut candidates = Vec::new();
for hist_layer in layers.iter_historic_layers() {
if !self.layer_cache.contains(&hist_layer.filename()) {
if hist_layer.is_remote_layer() {
continue;
}

View File

@@ -17,7 +17,7 @@ use tokio::io::{AsyncRead, AsyncWrite};
use tokio_util::sync::CancellationToken;
use utils::{project_git_version, sentry_init::init_sentry};
use tracing::{error, info, warn};
use tracing::{error, info, warn, Instrument};
project_git_version!(GIT_VERSION);
@@ -141,7 +141,6 @@ async fn task_main(
tokio::select! {
accept_result = listener.accept() => {
let (socket, peer_addr) = accept_result?;
info!("accepted postgres client connection from {peer_addr}");
let session_id = uuid::Uuid::new_v4();
let tls_config = Arc::clone(&tls_config);
@@ -149,18 +148,18 @@ async fn task_main(
connections.spawn(
async move {
info!("spawned a task for {peer_addr}");
socket
.set_nodelay(true)
.context("failed to set socket option")?;
handle_client(dest_suffix, tls_config, session_id, socket).await
info!(%peer_addr, "serving");
handle_client(dest_suffix, tls_config, socket).await
}
.unwrap_or_else(|e| {
// Acknowledge that the task has finished with an error.
error!("per-client task finished with an error: {e:#}");
}),
})
.instrument(tracing::info_span!("handle_client", ?session_id))
);
}
_ = cancellation_token.cancelled() => {
@@ -192,7 +191,6 @@ async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
let mut stream = PqStream::new(Stream::from_raw(raw_stream));
let msg = stream.read_startup_packet().await?;
info!("received {msg:?}");
use pq_proto::FeStartupPacket::*;
match msg {
@@ -215,15 +213,19 @@ async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
}
Ok(raw.upgrade(tls_config).await?)
}
_ => stream.throw_error_str(ERR_INSECURE_CONNECTION).await?,
unexpected => {
info!(
?unexpected,
"unexpected startup packet, rejecting connection"
);
stream.throw_error_str(ERR_INSECURE_CONNECTION).await?
}
}
}
#[tracing::instrument(fields(session_id = ?session_id), skip_all)]
async fn handle_client(
dest_suffix: Arc<String>,
tls_config: Arc<rustls::ServerConfig>,
session_id: uuid::Uuid,
stream: impl AsyncRead + AsyncWrite + Unpin,
) -> anyhow::Result<()> {
let tls_stream = ssl_handshake(stream, tls_config).await?;

View File

@@ -65,12 +65,19 @@ PAGESERVER_PER_TENANT_METRICS: Tuple[str, ...] = (
"pageserver_getpage_reconstruct_seconds_bucket",
"pageserver_getpage_reconstruct_seconds_count",
"pageserver_getpage_reconstruct_seconds_sum",
"pageserver_getpage_get_reconstruct_data_seconds_bucket",
"pageserver_getpage_get_reconstruct_data_seconds_count",
"pageserver_getpage_get_reconstruct_data_seconds_sum",
"pageserver_io_operations_bytes_total",
"pageserver_io_operations_seconds_bucket",
"pageserver_io_operations_seconds_count",
"pageserver_io_operations_seconds_sum",
"pageserver_last_record_lsn",
"pageserver_materialized_cache_hits_total",
"pageserver_materialized_cache_hits_direct_total",
"pageserver_read_num_fs_layers_bucket",
"pageserver_read_num_fs_layers_count",
"pageserver_read_num_fs_layers_sum",
"pageserver_smgr_query_seconds_bucket",
"pageserver_smgr_query_seconds_count",
"pageserver_smgr_query_seconds_sum",

View File

@@ -158,6 +158,7 @@ def test_fully_custom_config(positive_env: NeonEnv):
"threshold": "23h",
},
"evictions_low_residence_duration_metric_threshold": "2days",
"gc_feedback": True,
"gc_horizon": 23 * (1024 * 1024),
"gc_period": "2h 13m",
"image_creation_threshold": 7,

View File

@@ -37,6 +37,7 @@ class PgSniRouter(PgProtocol):
destination: str,
tls_cert: Path,
tls_key: Path,
test_output_dir: Path,
):
# Must use a hostname rather than IP here, for SNI to work
host = "localhost"
@@ -49,6 +50,7 @@ class PgSniRouter(PgProtocol):
self.tls_cert = tls_cert
self.tls_key = tls_key
self._popen: Optional[subprocess.Popen[bytes]] = None
self.test_output_dir = test_output_dir
def start(self) -> "PgSniRouter":
assert self._popen is None
@@ -60,8 +62,12 @@ class PgSniRouter(PgProtocol):
*["--destination", self.destination],
]
self._popen = subprocess.Popen(args)
router_log_path = self.test_output_dir / "pg_sni_router.log"
router_log = open(router_log_path, "w")
self._popen = subprocess.Popen(args, stderr=router_log)
self._wait_until_ready()
log.info(f"pg_sni_router started, log file: {router_log_path}")
return self
@backoff.on_exception(backoff.expo, OSError, max_time=10)
@@ -121,6 +127,7 @@ def test_pg_sni_router(
destination="localtest.me",
tls_cert=test_output_dir / "router.crt",
tls_key=test_output_dir / "router.key",
test_output_dir=test_output_dir,
) as router:
router.start()

View File

@@ -437,12 +437,22 @@ def test_delete_timeline_client_hangup(neon_env_builder: NeonEnvBuilder):
wait_until(50, 0.1, got_hangup_log_message)
# ok, retry without failpoint, it should succeed
# check that the timeline is still present
ps_http.timeline_detail(env.initial_tenant, child_timeline_id)
# ok, disable the failpoint to let the deletion finish
ps_http.configure_failpoints((failpoint_name, "off"))
# this should succeed
ps_http.timeline_delete(env.initial_tenant, child_timeline_id, timeout=2)
# the second call will try to transition the timeline into Stopping state, but it's already in that state
env.pageserver.allowed_errors.append(
f".*{child_timeline_id}.*Ignoring new state, equal to the existing one: Stopping"
)
def first_request_finished():
message = f".*DELETE.*{child_timeline_id}.*Cancelled request finished"
assert env.pageserver.log_contains(message)
wait_until(50, 0.1, first_request_finished)
# check that the timeline is gone
notfound_message = f"Timeline {env.initial_tenant}/{child_timeline_id} was not found"
env.pageserver.allowed_errors.append(".*" + notfound_message)
with pytest.raises(PageserverApiException, match=notfound_message) as exc:
ps_http.timeline_detail(env.initial_tenant, child_timeline_id)
assert exc.value.status_code == 404