Compare commits

...

4 Commits

Author SHA1 Message Date
Gleb Novikov
716851daa1 Set LD_LIBRARY_PATH in compute-node.Dockerfile 2024-12-20 11:55:48 +00:00
Alex Chi Z.
9c53b41245 fix(pageserver): update remote latest_gc_cutoff after gc-compaction (#10209)
## Problem

close https://github.com/neondatabase/neon/issues/10208
part of #9114 

## Summary of changes

* Ensure remote `latest_gc_cutoff` is up-to-date before removing any
files for gc-compaction.

Signed-off-by: Alex Chi Z <chi@neon.tech>
2024-12-19 18:40:20 +00:00
Konstantin Knizhnik
197a89ab3d Increase default stotrage controller heartbeat interval from 100msec … (#10206)
## Problem

Currently default value of storage controller heartbeat interval is
100msec. It means that 10 times per second it establish connection to
PS. And it seems to be quite expensive.
At MacOS right now storage_controller consumes 70% CPU and trusts - 30%.
So together they completely utilize one core.
A lot of us has Macs. Let's save environment a little bit and do not
waste electricity and contribute to global warming.

By the way, on prod we have interval  10seconds 

## Summary of changes

Increase heartbeat interval from 100msec to 1 second.

Co-authored-by: Konstantin Knizhnik <knizhnik@neon.tech>
2024-12-19 18:32:32 +00:00
Alex Chi Z.
b89e02f3e8 fix(pageserver): consider partial compaction layer map in layer check (#10044)
## Problem

In https://github.com/neondatabase/neon/pull/9897 we temporarily
disabled the layer valid check because the current one only considers
the end result of all compaction algorithms, but partial gc-compaction
would temporarily produce an "invalid" layer map.

part of https://github.com/neondatabase/neon/issues/9114

## Summary of changes

Allow LSN splits to overlap in the slow path check. Currently, the valid
check is only used in storage scrubber (background job) and during
gc-compaction (without taking layer lock). Therefore, it's fine for such
checks to be a little bit inefficient but more accurate.

---------

Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
2024-12-19 18:04:53 +00:00
4 changed files with 80 additions and 25 deletions

View File

@@ -1583,4 +1583,5 @@ RUN set -ex; \
ENV LANG=en_US.utf8
USER postgres
ENV LD_LIBRARY_PATH="/usr/local/lib/:$LD_LIBRARY_PATH"
ENTRYPOINT ["/usr/local/bin/compute_ctl"]

View File

@@ -180,7 +180,7 @@ impl NeonStorageControllerConf {
const DEFAULT_MAX_WARMING_UP_INTERVAL: std::time::Duration = std::time::Duration::from_secs(30);
// Very tight heartbeat interval to speed up tests
const DEFAULT_HEARTBEAT_INTERVAL: std::time::Duration = std::time::Duration::from_millis(100);
const DEFAULT_HEARTBEAT_INTERVAL: std::time::Duration = std::time::Duration::from_millis(1000);
}
impl Default for NeonStorageControllerConf {

View File

@@ -1,12 +1,15 @@
use std::collections::BTreeSet;
use itertools::Itertools;
use pageserver_compaction::helpers::overlaps_with;
use super::storage_layer::LayerName;
/// Checks whether a layer map is valid (i.e., is a valid result of the current compaction algorithm if nothing goes wrong).
///
/// The function checks if we can split the LSN range of a delta layer only at the LSNs of the delta layers. For example,
/// The function implements a fast path check and a slow path check.
///
/// The fast path checks if we can split the LSN range of a delta layer only at the LSNs of the delta layers. For example,
///
/// ```plain
/// | | | |
@@ -25,31 +28,47 @@ use super::storage_layer::LayerName;
/// | | | 4 | | |
///
/// If layer 2 and 4 contain the same single key, this is also a valid layer map.
///
/// However, if a partial compaction is still going on, it is possible that we get a layer map not satisfying the above condition.
/// Therefore, we fallback to simply check if any of the two delta layers overlap. (See "A slow path...")
pub fn check_valid_layermap(metadata: &[LayerName]) -> Option<String> {
let mut lsn_split_point = BTreeSet::new(); // TODO: use a better data structure (range tree / range set?)
let mut all_delta_layers = Vec::new();
for name in metadata {
if let LayerName::Delta(layer) = name {
if layer.key_range.start.next() != layer.key_range.end {
all_delta_layers.push(layer.clone());
}
all_delta_layers.push(layer.clone());
}
}
for layer in &all_delta_layers {
let lsn_range = &layer.lsn_range;
lsn_split_point.insert(lsn_range.start);
lsn_split_point.insert(lsn_range.end);
if layer.key_range.start.next() != layer.key_range.end {
let lsn_range = &layer.lsn_range;
lsn_split_point.insert(lsn_range.start);
lsn_split_point.insert(lsn_range.end);
}
}
for layer in &all_delta_layers {
for (idx, layer) in all_delta_layers.iter().enumerate() {
if layer.key_range.start.next() == layer.key_range.end {
continue;
}
let lsn_range = layer.lsn_range.clone();
let intersects = lsn_split_point.range(lsn_range).collect_vec();
if intersects.len() > 1 {
let err = format!(
"layer violates the layer map LSN split assumption: layer {} intersects with LSN [{}]",
layer,
intersects.into_iter().map(|lsn| lsn.to_string()).join(", ")
);
return Some(err);
// A slow path to check if the layer intersects with any other delta layer.
for (other_idx, other_layer) in all_delta_layers.iter().enumerate() {
if other_idx == idx {
// do not check self intersects with self
continue;
}
if overlaps_with(&layer.lsn_range, &other_layer.lsn_range)
&& overlaps_with(&layer.key_range, &other_layer.key_range)
{
let err = format!(
"layer violates the layer map LSN split assumption: layer {} intersects with layer {}",
layer, other_layer
);
return Some(err);
}
}
}
}
None

View File

@@ -29,6 +29,7 @@ use utils::id::TimelineId;
use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
use crate::page_cache;
use crate::statvfs::Statvfs;
use crate::tenant::checks::check_valid_layermap;
use crate::tenant::remote_timeline_client::WaitCompletionError;
use crate::tenant::storage_layer::batch_split_writer::{
BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
@@ -2156,15 +2157,14 @@ impl Timeline {
// Step 1: construct a k-merge iterator over all layers.
// Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
// disable the check for now because we need to adjust the check for partial compactions, will enable later.
// let layer_names = job_desc
// .selected_layers
// .iter()
// .map(|layer| layer.layer_desc().layer_name())
// .collect_vec();
// if let Some(err) = check_valid_layermap(&layer_names) {
// warn!("gc-compaction layer map check failed because {}, this is normal if partial compaction is not finished yet", err);
// }
let layer_names = job_desc
.selected_layers
.iter()
.map(|layer| layer.layer_desc().layer_name())
.collect_vec();
if let Some(err) = check_valid_layermap(&layer_names) {
bail!("gc-compaction layer map check failed because {}, cannot proceed with compaction due to potential data loss", err);
}
// The maximum LSN we are processing in this compaction loop
let end_lsn = job_desc
.selected_layers
@@ -2546,13 +2546,48 @@ impl Timeline {
);
// Step 3: Place back to the layer map.
// First, do a sanity check to ensure the newly-created layer map does not contain overlaps.
let all_layers = {
let guard = self.layers.read().await;
let layer_map = guard.layer_map()?;
layer_map.iter_historic_layers().collect_vec()
};
let mut final_layers = all_layers
.iter()
.map(|layer| layer.layer_name())
.collect::<HashSet<_>>();
for layer in &layer_selection {
final_layers.remove(&layer.layer_desc().layer_name());
}
for layer in &compact_to {
final_layers.insert(layer.layer_desc().layer_name());
}
let final_layers = final_layers.into_iter().collect_vec();
// TODO: move this check before we call `finish` on image layer writers. However, this will require us to get the layer name before we finish
// the writer, so potentially, we will need a function like `ImageLayerBatchWriter::get_all_pending_layer_keys` to get all the keys that are
// in the writer before finalizing the persistent layers. Now we would leave some dangling layers on the disk if the check fails.
if let Some(err) = check_valid_layermap(&final_layers) {
bail!("gc-compaction layer map check failed after compaction because {}, compaction result not applied to the layer map due to potential data loss", err);
}
// Between the sanity check and this compaction update, there could be new layers being flushed, but it should be fine because we only
// operate on L1 layers.
{
// TODO: sanity check if the layer map is valid (i.e., should not have overlaps)
let mut guard = self.layers.write().await;
guard
.open_mut()?
.finish_gc_compaction(&layer_selection, &compact_to, &self.metrics)
};
// Schedule an index-only upload to update the `latest_gc_cutoff` in the index_part.json.
// Otherwise, after restart, the index_part only contains the old `latest_gc_cutoff` and
// find_gc_cutoffs will try accessing things below the cutoff. TODO: ideally, this should
// be batched into `schedule_compaction_update`.
let disk_consistent_lsn = self.disk_consistent_lsn.load();
self.schedule_uploads(disk_consistent_lsn, None)?;
self.remote_client
.schedule_compaction_update(&layer_selection, &compact_to)?;