pageserver: implementation of update_layer_visibility

This commit is contained in:
John Spray
2024-06-29 20:54:49 +01:00
parent c78c810118
commit d7bca9fcdb
6 changed files with 270 additions and 3 deletions

View File

@@ -500,6 +500,15 @@ static RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
.expect("failed to define a metric")
});
static VISIBLE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
register_uint_gauge_vec!(
"pageserver_visible_physical_size",
"The size of the layer files present in the pageserver's filesystem.",
&["tenant_id", "shard_id", "timeline_id"]
)
.expect("failed to define a metric")
});
pub(crate) static RESIDENT_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
register_uint_gauge!(
"pageserver_resident_physical_size_global",
@@ -2130,6 +2139,7 @@ pub(crate) struct TimelineMetrics {
pub archival_size: UIntGauge,
pub standby_horizon_gauge: IntGauge,
pub resident_physical_size_gauge: UIntGauge,
pub visible_physical_size_gauge: UIntGauge,
/// copy of LayeredTimeline.current_logical_size
pub current_logical_size_gauge: UIntGauge,
pub aux_file_size_gauge: IntGauge,
@@ -2216,6 +2226,9 @@ impl TimelineMetrics {
let resident_physical_size_gauge = RESIDENT_PHYSICAL_SIZE
.get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
.unwrap();
let visible_physical_size_gauge = VISIBLE_PHYSICAL_SIZE
.get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
.unwrap();
// TODO: we shouldn't expose this metric
let current_logical_size_gauge = CURRENT_LOGICAL_SIZE
.get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
@@ -2266,6 +2279,7 @@ impl TimelineMetrics {
archival_size,
standby_horizon_gauge,
resident_physical_size_gauge,
visible_physical_size_gauge,
current_logical_size_gauge,
aux_file_size_gauge,
directory_entries_count_gauge,
@@ -2317,6 +2331,7 @@ impl TimelineMetrics {
RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
}
let _ = VISIBLE_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
if let Some(metric) = Lazy::get(&DIRECTORY_ENTRIES_COUNT) {
let _ = metric.remove_label_values(&[tenant_id, shard_id, timeline_id]);

View File

@@ -51,7 +51,7 @@ use crate::keyspace::KeyPartitioning;
use crate::repository::Key;
use crate::tenant::storage_layer::InMemoryLayer;
use anyhow::Result;
use pageserver_api::keyspace::KeySpaceAccum;
use pageserver_api::keyspace::{KeySpace, KeySpaceAccum, KeySpaceRandomAccum};
use std::collections::{HashMap, VecDeque};
use std::iter::Peekable;
use std::ops::Range;
@@ -61,7 +61,7 @@ use utils::lsn::Lsn;
use historic_layer_coverage::BufferedHistoricLayerCoverage;
pub use historic_layer_coverage::LayerKey;
use super::storage_layer::PersistentLayerDesc;
use super::storage_layer::{LayerVisibility, PersistentLayerDesc};
///
/// LayerMap tracks what layers exist on a timeline.
@@ -870,6 +870,164 @@ impl LayerMap {
println!("End dump LayerMap");
Ok(())
}
/// `read_points` represent the tip of a timeline and any branch points, i.e. the places
/// where we expect to serve reads.
///
/// This function is O(N) and should be called infrequently. The caller is responsible for
/// looking up and updating the Layer objects for these layer descriptors.
pub(crate) fn get_visibility(
&self,
mut read_points: Vec<(Lsn, KeySpace)>,
) -> (Vec<(Arc<PersistentLayerDesc>, LayerVisibility)>, KeySpace) {
// This is like a KeySpace, but written for efficient subtraction of layers and unions with KeySpaces
struct KeyShadow {
// FIXME: consider efficiency. KeySpace is a flat vector, so in principle fairly inefficient for
// repeatedly calling contains(), BUT as we iterate through the layermap we expect the shadow to shrink
// to something quite small, and for small collections an algorithmically expensive vector is often better
// for performance than a more algorithmically cheap data structure.
inner: KeySpace,
}
impl KeyShadow {
fn new(keyspace: KeySpace) -> Self {
Self { inner: keyspace }
}
fn contains(&self, range: Range<Key>) -> bool {
self.inner.overlaps(&range)
}
/// Return true if anything was removed.
fn subtract(&mut self, range: Range<Key>) -> bool {
let removed = self.inner.remove_overlapping_with(&KeySpace {
ranges: vec![range],
});
!removed.ranges.is_empty()
}
fn union_with(&mut self, keyspace: KeySpace) {
let mut accum = KeySpaceRandomAccum::new();
let prev = std::mem::take(&mut self.inner);
accum.add_keyspace(prev);
accum.add_keyspace(keyspace);
self.inner = accum.to_keyspace();
}
}
// The 'shadow' will be updated as we sweep through the layers: an image layer subtracts from the shadow,
// and a ReadPoint
read_points.sort_by_key(|rp| rp.0);
let mut shadow = KeyShadow::new(
read_points
.pop()
.expect("Every timeline has at least one read point")
.1,
);
// We will interleave all our read points and layers into a sorted collection
enum Item {
ReadPoint { lsn: Lsn, keyspace: KeySpace },
Layer(Arc<PersistentLayerDesc>),
}
let mut items = Vec::with_capacity(self.historic.len() + read_points.len());
items.extend(self.iter_historic_layers().map(Item::Layer));
items.extend(read_points.into_iter().map(|rp| Item::ReadPoint {
lsn: rp.0,
keyspace: rp.1,
}));
// Ordering: we want to iterate like this:
// 1. Highest LSNs first
// 2. Consider ReadPoints before image layers if they're at the same LSN
items.sort_by_key(|item| {
std::cmp::Reverse(match item {
Item::ReadPoint {
lsn,
keyspace: _keyspace,
} => (*lsn, 0),
Item::Layer(layer) => {
if layer.is_delta() {
(layer.get_lsn_range().end, 1)
} else {
(layer.image_layer_lsn(), 2)
}
}
})
});
let mut results = Vec::with_capacity(self.historic.len());
// TODO: handle delta layers properly with multiple read points: if a read point intersects a delta layer, we might already
// have encountered it and marked it as not-visible. We need to keep track of which delta layers we are currently within, and
// when we encounter a ReadPoint, update the delta layer's visibility as needed.
// let mut pending_delta : Vec= ...
let mut maybe_covered_deltas: Vec<Arc<PersistentLayerDesc>> = Vec::new();
for item in items {
let (reached_lsn, is_readpoint) = match &item {
Item::ReadPoint {
lsn,
keyspace: _keyspace,
} => (lsn, true),
Item::Layer(layer) => (&layer.lsn_range.start, false),
};
maybe_covered_deltas.retain(|d| {
if *reached_lsn >= d.lsn_range.start && is_readpoint {
// We encountered a readpoint within the delta layer: it is visible
results.push((d.clone(), LayerVisibility::Visible));
false
} else if *reached_lsn < d.lsn_range.start {
// We passed the layer's range without encountering a read point: it is not visible
results.push((d.clone(), LayerVisibility::Covered));
false
} else {
// We're still in the delta layer: continue iterating
true
}
});
match item {
Item::ReadPoint {
lsn: _lsn,
keyspace,
} => {
shadow.union_with(keyspace);
}
Item::Layer(layer) => {
let visibility = if layer.is_delta() {
if shadow.contains(layer.get_key_range()) {
LayerVisibility::Visible
} else {
// If a layer isn't visible based on current state, we must defer deciding whether
// it is truly not visible until we have advanced past the delta's range: we might
// encounter another branch point within this delta layer's LSN range.
maybe_covered_deltas.push(layer);
continue;
}
} else if shadow.subtract(layer.get_key_range()) {
// An image layer, which overlapped with the shadow
LayerVisibility::Visible
} else {
// An image layer, which did not overlap with the shadow
LayerVisibility::Covered
};
results.push((layer, visibility));
}
}
}
// Drain any remaining maybe_covered deltas
results.extend(
maybe_covered_deltas
.into_iter()
.map(|d| (d, LayerVisibility::Covered)),
);
(results, shadow.inner)
}
}
#[cfg(test)]

View File

@@ -521,6 +521,10 @@ impl<Value: Clone> BufferedHistoricLayerCoverage<Value> {
Ok(&self.historic_coverage)
}
pub(crate) fn len(&self) -> usize {
self.layers.len()
}
}
#[test]

View File

@@ -700,6 +700,10 @@ impl LayerAccessStats {
},
}
}
pub(crate) fn set_visibility(&self, visibility: LayerVisibility) {
self.0.lock().unwrap().visibility = visibility;
}
}
/// Get a layer descriptor from a layer.

View File

@@ -1817,6 +1817,10 @@ impl Timeline {
self.compact_shard_ancestors(rewrite_max, ctx).await?;
}
// TODO: be more selective: call this once at startup, and thereafter only when some branching changes or
// when image layer are generated.
self.update_layer_visibility(ctx).await?;
Ok(())
}

View File

@@ -26,7 +26,7 @@ use utils::id::TimelineId;
use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
use crate::page_cache;
use crate::tenant::storage_layer::{AsLayerDesc, PersistentLayerDesc};
use crate::tenant::storage_layer::{AsLayerDesc, LayerVisibility, PersistentLayerDesc};
use crate::tenant::timeline::{drop_rlock, Hole, ImageLayerCreationOutcome};
use crate::tenant::timeline::{DeltaLayerWriter, ImageLayerWriter};
use crate::tenant::timeline::{Layer, ResidentLayer};
@@ -347,6 +347,88 @@ impl Timeline {
Ok(())
}
/// A post-compaction step to update the LayerVisibility of layers covered by image layers. This
/// should also be called when new branches are created.
///
/// Sweep through the layer map, identifying layers which are covered by image layers
/// such that they do not need to be available to service reads. The resulting LayerVisibility
/// result may be used as an input to eviction and secondary downloads to de-prioritize layers
/// that we know won't be needed for reads.
pub(super) async fn update_layer_visibility(
&self,
ctx: &RequestContext,
) -> Result<(), CompactionError> {
// Start with a keyspace representing all the keys we need to read from the tip of the branch
let head_lsn = self.get_last_record_lsn();
let (mut head_keyspace, sparse_ks) = self.collect_keyspace(head_lsn, ctx).await?;
// Converting the sparse part of the keyspace into the dense keyspace is safe in this context
// because we will never iterate through the keys.
head_keyspace.merge(&sparse_ks.0);
// We will sweep through layers in reverse-LSN order. We only do historic layers. L0 deltas
// are implicitly visible, because LayerVisibility's default is Visible, and we never modify it here.
let layer_manager = self.layers.read().await;
let layer_map = layer_manager.layer_map();
let mut visible_size: u64 = 0;
// FIXME: we only get accurate keyspaces from children if they've already run update_layer_visibility themselves. At startup all the timelines
// initialize this in arbitrary order (at the end of initial_logical_size_calculation). We should coordinate these. Perhaps at the very start
// of the tenant compaction task we should do all the timelines' layer visibility calculations in a leaf-first order?
let readable_points = {
let children = self.gc_info.read().unwrap().retain_lsns.clone();
let mut readable_points = Vec::with_capacity(children.len() + 1);
for (child_lsn, _child_timeline_id, child_keyspace) in &children {
let keyspace = match child_keyspace {
Some(ks) => ks.clone(),
None => {
// The child has not posted information about which parts of the keyspace they depend on: presume they depend on all of it.
let (mut keyspace, sparse_keyspace) =
self.collect_keyspace(*child_lsn, ctx).await?;
keyspace.merge(&sparse_keyspace.0);
keyspace
}
};
readable_points.push((*child_lsn, keyspace));
}
readable_points.push((head_lsn, head_keyspace));
readable_points
};
let (layer_visibility, shadow) = layer_map.get_visibility(readable_points);
for (layer_desc, visibility) in layer_visibility {
// FIXME: a more efficiency bulk zip() through the layers rather than NlogN getting each one
let layer = layer_manager.get_from_desc(&layer_desc);
if matches!(visibility, LayerVisibility::Visible) {
visible_size += layer.metadata().file_size;
}
layer.access_stats().set_visibility(visibility);
}
if let Some(ancestor) = &self.ancestor_timeline {
// Having calculated the readable keyspace after walking back through all this timeline's layers, the resulting keyspace is the remaining
// keys for which reads may still fall through to the parent branch. Notify the parent branch of this, so that they may GC layers which
// do not overlap with this keyspace, and so that they may use this as an input to their own visibility updates.
ancestor
.gc_info
.write()
.unwrap()
.notify_child_keyspace(self.timeline_id, shadow);
}
// Also include in the visible size all the layers which we would never update visibility on
// TODO: getter that doesn't spuriously construct a Vec<>
for layer in layer_map.get_level0_deltas().unwrap() {
visible_size += layer.file_size;
}
self.metrics.visible_physical_size_gauge.set(visible_size);
Ok(())
}
/// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
/// as Level 1 files.
async fn compact_level0(