Compare commits

..

7 Commits

Author SHA1 Message Date
Bojan Serafimov ce3cf7ed56 wip 2023-11-03 16:41:57 -04:00
Bojan Serafimov ebbc71076b Merge branch 'fast-sort' of github.com:neondatabase/neon into fast-sort 2023-11-03 14:44:58 -04:00
Bojan Serafimov 0b124f882f Fix field1 input in test data 2023-11-03 14:44:40 -04:00
bojanserafimov b3b0eda846 Merge branch 'main' into fast-sort 2023-11-02 09:58:06 -04:00
Bojan Serafimov fe5f79aed6 unstable sort 2023-11-02 09:57:23 -04:00
Bojan Serafimov a4047fb9de Use valid keys in tests 2023-11-02 09:54:27 -04:00
Bojan Serafimov 622e7f07b4 Speed up sort 2023-11-01 22:38:19 -04:00
17 changed files with 2148 additions and 2131 deletions
+2 -2
View File
@@ -283,6 +283,7 @@ fn main() -> Result<()> {
.expect("--vm-monitor-addr should always be set because it has a default arg");
let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
let cgroup = matches.get_one::<String>("cgroup");
let file_cache_on_disk = matches.get_flag("file-cache-on-disk");
// Only make a runtime if we need to.
// Note: it seems like you can make a runtime in an inner scope and
@@ -309,6 +310,7 @@ fn main() -> Result<()> {
cgroup: cgroup.cloned(),
pgconnstr: file_cache_connstr.cloned(),
addr: vm_monitor_addr.clone(),
file_cache_on_disk,
})),
token.clone(),
))
@@ -480,8 +482,6 @@ fn cli() -> clap::Command {
.value_name("FILECACHE_CONNSTR"),
)
.arg(
// DEPRECATED, NO LONGER DOES ANYTHING.
// See https://github.com/neondatabase/cloud/issues/7516
Arg::new("file-cache-on-disk")
.long("file-cache-on-disk")
.action(clap::ArgAction::SetTrue),
+4 -37
View File
@@ -1,7 +1,4 @@
use std::sync::{
atomic::{AtomicUsize, Ordering},
Arc, Mutex, MutexGuard,
};
use std::sync::{Arc, Mutex, MutexGuard};
use tokio::sync::Semaphore;
/// Custom design like [`tokio::sync::OnceCell`] but using [`OwnedSemaphorePermit`] instead of
@@ -13,7 +10,6 @@ use tokio::sync::Semaphore;
/// [`OwnedSemaphorePermit`]: tokio::sync::OwnedSemaphorePermit
pub struct OnceCell<T> {
inner: Mutex<Inner<T>>,
initializers: AtomicUsize,
}
impl<T> Default for OnceCell<T> {
@@ -21,7 +17,6 @@ impl<T> Default for OnceCell<T> {
fn default() -> Self {
Self {
inner: Default::default(),
initializers: AtomicUsize::new(0),
}
}
}
@@ -54,7 +49,6 @@ impl<T> OnceCell<T> {
init_semaphore: Arc::new(sem),
value: Some(value),
}),
initializers: AtomicUsize::new(0),
}
}
@@ -77,11 +71,7 @@ impl<T> OnceCell<T> {
guard.init_semaphore.clone()
};
let permit = {
// increment the count for the duration of queued
let _guard = CountWaitingInitializers::start(self);
sem.acquire_owned().await
};
let permit = sem.acquire_owned().await;
match permit {
Ok(permit) => {
@@ -110,13 +100,12 @@ impl<T> OnceCell<T> {
///
/// If the inner has already been initialized.
pub fn set(&self, value: T, _permit: InitPermit) -> Guard<'_, T> {
// cannot assert that this permit is for self.inner.semaphore
let guard = self.inner.lock().unwrap();
// cannot assert that this permit is for self.inner.semaphore, but we can assert it cannot
// give more permits right now.
if guard.init_semaphore.try_acquire().is_ok() {
drop(guard);
panic!("permit is of wrong origin");
panic!("semaphore is of wrong origin");
}
Self::set0(value, guard)
@@ -141,28 +130,6 @@ impl<T> OnceCell<T> {
None
}
}
/// Return the number of [`Self::get_or_init`] calls waiting for initialization to complete.
pub fn initializer_count(&self) -> usize {
self.initializers.load(Ordering::Relaxed)
}
}
/// DropGuard counter for queued tasks waiting to initialize, mainly accessible for the
/// initializing task for example at the end of initialization.
struct CountWaitingInitializers<'a, T>(&'a OnceCell<T>);
impl<'a, T> CountWaitingInitializers<'a, T> {
fn start(target: &'a OnceCell<T>) -> Self {
target.initializers.fetch_add(1, Ordering::Relaxed);
CountWaitingInitializers(target)
}
}
impl<'a, T> Drop for CountWaitingInitializers<'a, T> {
fn drop(&mut self) {
self.0.initializers.fetch_sub(1, Ordering::Relaxed);
}
}
/// Uninteresting guard object to allow short-lived access to inspect or clone the held,
+20 -4
View File
@@ -21,6 +21,11 @@ pub struct FileCacheState {
#[derive(Debug)]
pub struct FileCacheConfig {
/// Whether the file cache is *actually* stored in memory (e.g. by writing to
/// a tmpfs or shmem file). If true, the size of the file cache will be counted against the
/// memory available for the cgroup.
pub(crate) in_memory: bool,
/// The size of the file cache, in terms of the size of the resource it consumes
/// (currently: only memory)
///
@@ -54,9 +59,22 @@ pub struct FileCacheConfig {
spread_factor: f64,
}
impl Default for FileCacheConfig {
fn default() -> Self {
impl FileCacheConfig {
pub fn default_in_memory() -> Self {
Self {
in_memory: true,
// 75 %
resource_multiplier: 0.75,
// 640 MiB; (512 + 128)
min_remaining_after_cache: NonZeroU64::new(640 * MiB).unwrap(),
// ensure any increase in file cache size is split 90-10 with 10% to other memory
spread_factor: 0.1,
}
}
pub fn default_on_disk() -> Self {
Self {
in_memory: false,
resource_multiplier: 0.75,
// 256 MiB - lower than when in memory because overcommitting is safe; if we don't have
// memory, the kernel will just evict from its page cache, rather than e.g. killing
@@ -65,9 +83,7 @@ impl Default for FileCacheConfig {
spread_factor: 0.1,
}
}
}
impl FileCacheConfig {
/// Make sure fields of the config are consistent.
pub fn validate(&self) -> anyhow::Result<()> {
// Single field validity
+10
View File
@@ -39,6 +39,16 @@ pub struct Args {
#[arg(short, long)]
pub pgconnstr: Option<String>,
/// Flag to signal that the Postgres file cache is on disk (i.e. not in memory aside from the
/// kernel's page cache), and therefore should not count against available memory.
//
// NB: Ideally this flag would directly refer to whether the file cache is in memory (rather
// than a roundabout way, via whether it's on disk), but in order to be backwards compatible
// during the switch away from an in-memory file cache, we had to default to the previous
// behavior.
#[arg(long)]
pub file_cache_on_disk: bool,
/// The address we should listen on for connection requests. For the
/// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
#[arg(short, long)]
+26 -9
View File
@@ -156,7 +156,10 @@ impl Runner {
// memory limits.
if let Some(connstr) = &args.pgconnstr {
info!("initializing file cache");
let config = FileCacheConfig::default();
let config = match args.file_cache_on_disk {
true => FileCacheConfig::default_on_disk(),
false => FileCacheConfig::default_in_memory(),
};
let mut file_cache = FileCacheState::new(connstr, config, token.clone())
.await
@@ -184,7 +187,10 @@ impl Runner {
info!("file cache size actually got set to {actual_size}")
}
file_cache_disk_size = actual_size;
if args.file_cache_on_disk {
file_cache_disk_size = actual_size;
}
state.filecache = Some(file_cache);
}
@@ -233,11 +239,17 @@ impl Runner {
let requested_mem = target.mem;
let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
let expected_file_cache_size = self
let (expected_file_cache_size, expected_file_cache_disk_size) = self
.filecache
.as_ref()
.map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
.unwrap_or(0);
.map(|file_cache| {
let size = file_cache.config.calculate_cache_size(usable_system_memory);
match file_cache.config.in_memory {
true => (size, 0),
false => (size, size),
}
})
.unwrap_or((0, 0));
if let Some(cgroup) = &self.cgroup {
let (last_time, last_history) = *cgroup.watcher.borrow();
@@ -261,7 +273,7 @@ impl Runner {
let new_threshold = self
.config
.cgroup_threshold(usable_system_memory, expected_file_cache_size);
.cgroup_threshold(usable_system_memory, expected_file_cache_disk_size);
let current = last_history.avg_non_reclaimable;
@@ -288,10 +300,13 @@ impl Runner {
.set_file_cache_size(expected_file_cache_size)
.await
.context("failed to set file cache size")?;
file_cache_disk_size = actual_usage;
if !file_cache.config.in_memory {
file_cache_disk_size = actual_usage;
}
let message = format!(
"set file cache size to {} MiB",
"set file cache size to {} MiB (in memory = {})",
bytes_to_mebibytes(actual_usage),
file_cache.config.in_memory,
);
info!("downscale: {message}");
status.push(message);
@@ -342,7 +357,9 @@ impl Runner {
.set_file_cache_size(expected_usage)
.await
.context("failed to set file cache size")?;
file_cache_disk_size = actual_usage;
if !file_cache.config.in_memory {
file_cache_disk_size = actual_usage;
}
if actual_usage != expected_usage {
warn!(
+5 -4
View File
@@ -28,8 +28,9 @@ impl Key {
/// As long as Neon does not support tablespace (because of lack of access to local file system),
/// we can assume that only some predefined namespace OIDs are used which can fit in u16
pub fn to_i128(&self) -> i128 {
assert!(self.field2 < 0xFFFF || self.field2 == 0xFFFFFFFF || self.field2 == 0x22222222);
(((self.field1 & 0xf) as i128) << 120)
assert!(self.field1 < 0xF);
assert!(self.field2 < 0xFFFF);
(((self.field1 & 0xF) as i128) << 120)
| (((self.field2 & 0xFFFF) as i128) << 104)
| ((self.field3 as i128) << 72)
| ((self.field4 as i128) << 40)
@@ -149,8 +150,8 @@ impl Key {
field6: u32::MIN,
};
pub const MAX: Key = Key {
field1: u8::MAX,
field2: u32::MAX,
field1: 0xF - 1,
field2: 0xFFFF - 1,
field3: u32::MAX,
field4: u32::MAX,
field5: u8::MAX,
+7 -8
View File
@@ -3340,7 +3340,6 @@ fn run_initdb(
.args(["-D", initdb_target_dir.as_ref()])
.args(["-U", &conf.superuser])
.args(["-E", "utf8"])
.args(["-l", "logical"])
.arg("--no-instructions")
// This is only used for a temporary installation that is deleted shortly after,
// so no need to fsync it
@@ -3693,7 +3692,7 @@ mod tests {
use tokio_util::sync::CancellationToken;
static TEST_KEY: Lazy<Key> =
Lazy::new(|| Key::from_slice(&hex!("112222222233333333444444445500000001")));
Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
#[tokio::test]
async fn test_basic() -> anyhow::Result<()> {
@@ -3789,9 +3788,9 @@ mod tests {
let writer = tline.writer().await;
#[allow(non_snake_case)]
let TEST_KEY_A: Key = Key::from_hex("112222222233333333444444445500000001").unwrap();
let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
#[allow(non_snake_case)]
let TEST_KEY_B: Key = Key::from_hex("112222222233333333444444445500000002").unwrap();
let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
// Insert a value on the timeline
writer
@@ -4375,7 +4374,7 @@ mod tests {
let mut keyspace = KeySpaceAccum::new();
let mut test_key = Key::from_hex("012222222233333333444444445500000000").unwrap();
let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
let mut blknum = 0;
for _ in 0..50 {
for _ in 0..10000 {
@@ -4421,7 +4420,7 @@ mod tests {
const NUM_KEYS: usize = 1000;
let mut test_key = Key::from_hex("012222222233333333444444445500000000").unwrap();
let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
let mut keyspace = KeySpaceAccum::new();
@@ -4502,7 +4501,7 @@ mod tests {
const NUM_KEYS: usize = 1000;
let mut test_key = Key::from_hex("012222222233333333444444445500000000").unwrap();
let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
let mut keyspace = KeySpaceAccum::new();
@@ -4593,7 +4592,7 @@ mod tests {
const NUM_KEYS: usize = 100;
const NUM_TLINES: usize = 50;
let mut test_key = Key::from_hex("012222222233333333444444445500000000").unwrap();
let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
// Track page mutation lsns across different timelines.
let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
File diff suppressed because it is too large Load Diff
@@ -345,14 +345,19 @@ impl InMemoryLayer {
let cursor = inner.file.block_cursor();
let mut keys: Vec<(&Key, &VecMap<Lsn, u64>)> = inner.index.iter().collect();
keys.sort_by_key(|k| k.0);
// Sort the keys because delta layer writer expects them sorted.
//
// NOTE: this sort can take up significant time if the layer has millions of
// keys. To speed up all the comparisons we convert the key to i128 and
// keep the value as a reference.
let mut keys: Vec<_> = inner.index.iter().map(|(k, m)| (k.to_i128(), m)).collect();
keys.sort_unstable_by_key(|k| k.0);
let ctx = RequestContextBuilder::extend(ctx)
.page_content_kind(PageContentKind::InMemoryLayer)
.build();
for (key, vec_map) in keys.iter() {
let key = **key;
let key = Key::from_i128(*key);
// Write all page versions
for (lsn, pos) in vec_map.as_slice() {
cursor.read_blob_into_buf(*pos, &mut buf, &ctx).await?;
+10 -30
View File
@@ -411,10 +411,6 @@ struct LayerInner {
version: AtomicUsize,
/// Allow subscribing to when the layer actually gets evicted.
///
/// If in future we need to implement "wait until layer instances are gone and done", carrying
/// this over to the gc spawn_blocking from LayerInner::drop will do the trick, and adding a
/// method for "wait_gc" which will wait to this being closed.
status: tokio::sync::broadcast::Sender<Status>,
/// Counter for exponential backoff with the download
@@ -565,8 +561,6 @@ impl LayerInner {
}
}
/// Cancellation safe, however dropping the future and calling this method again might result
/// in a new attempt to evict OR join the previously started attempt.
pub(crate) async fn evict_and_wait(
&self,
_: &RemoteTimelineClient,
@@ -615,8 +609,8 @@ impl LayerInner {
}
}
/// Cancellation safe.
#[tracing::instrument(skip_all, fields(layer=%self))]
/// Should be cancellation safe, but cancellation is troublesome together with the spawned
/// download.
async fn get_or_maybe_download(
self: &Arc<Self>,
allow_download: bool,
@@ -660,6 +654,8 @@ impl LayerInner {
return Err(DownloadError::NoRemoteStorage);
}
tracing::debug!(%reason, "downloading layer");
if let Some(ctx) = ctx {
self.check_expected_download(ctx)?;
}
@@ -670,8 +666,6 @@ impl LayerInner {
return Err(DownloadError::DownloadRequired);
}
tracing::info!(%reason, "downloading on-demand");
self.spawn_download_and_wait(timeline, permit).await?
} else {
// the file is present locally, probably by a previous but cancelled call to
@@ -692,11 +686,6 @@ impl LayerInner {
LayerResidenceEventReason::ResidenceChange,
);
let waiters = self.inner.initializer_count();
if waiters > 0 {
tracing::info!(waiters, "completing the on-demand download for other tasks");
}
Ok((ResidentOrWantedEvicted::Resident(res), permit))
};
@@ -757,8 +746,8 @@ impl LayerInner {
match b {
Download => Ok(()),
Warn | Error => {
tracing::info!(
"unexpectedly on-demand downloading for task kind {:?}",
tracing::warn!(
"unexpectedly on-demand downloading remote layer {self} for task kind {:?}",
ctx.task_kind()
);
crate::metrics::UNEXPECTED_ONDEMAND_DOWNLOADS.inc();
@@ -790,7 +779,6 @@ impl LayerInner {
// block tenant::mgr::remove_tenant_from_memory.
let this: Arc<Self> = self.clone();
crate::task_mgr::spawn(
&tokio::runtime::Handle::current(),
crate::task_mgr::TaskKind::RemoteDownloadTask,
@@ -799,7 +787,6 @@ impl LayerInner {
&task_name,
false,
async move {
let client = timeline
.remote_client
.as_ref()
@@ -861,7 +848,6 @@ impl LayerInner {
}
self.consecutive_failures.store(0, Ordering::Relaxed);
tracing::info!("on-demand download successful");
Ok(permit)
}
@@ -1054,14 +1040,11 @@ impl LayerInner {
Ok(())
}
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
tracing::error!(
layer_size = %self.desc.file_size,
"failed to evict layer from disk, it was already gone (metrics will be inaccurate)"
);
tracing::info!("failed to evict file from disk, it was already gone");
Err(EvictionCancelled::FileNotFound)
}
Err(e) => {
tracing::error!("failed to evict file from disk: {e:#}");
tracing::warn!("failed to evict file from disk: {e:#}");
Err(EvictionCancelled::RemoveFailed)
}
};
@@ -1141,8 +1124,6 @@ impl std::fmt::Display for NeedsDownload {
/// Existence of `DownloadedLayer` means that we have the file locally, and can later evict it.
pub(crate) struct DownloadedLayer {
owner: Weak<LayerInner>,
// Use tokio OnceCell as we do not need to deinitialize this, it'll just get dropped with the
// DownloadedLayer
kind: tokio::sync::OnceCell<anyhow::Result<LayerKind>>,
version: usize,
}
@@ -1186,6 +1167,7 @@ impl DownloadedLayer {
"these are the same, just avoiding the upgrade"
);
// there is nothing async here, but it should be async
let res = if owner.desc.is_delta {
let summary = Some(delta_layer::Summary::expected(
owner.desc.tenant_id,
@@ -1284,8 +1266,6 @@ impl std::fmt::Debug for ResidentLayer {
impl ResidentLayer {
/// Release the eviction guard, converting back into a plain [`Layer`].
///
/// You can access the [`Layer`] also by using `as_ref`.
pub(crate) fn drop_eviction_guard(self) -> Layer {
self.into()
}
@@ -1341,7 +1321,7 @@ impl AsRef<Layer> for ResidentLayer {
}
}
/// Drop the eviction guard.
/// Allow slimming down if we don't want the `2*usize` with eviction candidates?
impl From<ResidentLayer> for Layer {
fn from(value: ResidentLayer) -> Self {
value.owner
-28
View File
@@ -13,7 +13,6 @@ pub struct ConsoleError {
#[derive(Deserialize)]
pub struct GetRoleSecret {
pub role_secret: Box<str>,
pub allowed_ips: Option<Vec<Box<str>>>,
}
// Manually implement debug to omit sensitive info.
@@ -188,31 +187,4 @@ mod tests {
Ok(())
}
#[test]
fn parse_wake_compute() -> anyhow::Result<()> {
let json = json!({
"address": "0.0.0.0",
"aux": dummy_aux(),
});
let _: WakeCompute = serde_json::from_str(&json.to_string())?;
Ok(())
}
#[test]
fn parse_get_role_secret() -> anyhow::Result<()> {
// Empty `allowed_ips` field.
let json = json!({
"role_secret": "secret",
});
let _: GetRoleSecret = serde_json::from_str(&json.to_string())?;
// Empty `allowed_ips` field.
let json = json!({
"role_secret": "secret",
"allowed_ips": ["8.8.8.8"],
});
let _: GetRoleSecret = serde_json::from_str(&json.to_string())?;
Ok(())
}
}
+1
View File
@@ -81,6 +81,7 @@ FALLBACK_DURATION = {
"test_runner/performance/test_seqscans.py::test_seqscans[vanilla-100000-100-0]": 0.55,
"test_runner/performance/test_seqscans.py::test_seqscans[vanilla-10000000-1-0]": 12.189,
"test_runner/performance/test_seqscans.py::test_seqscans[vanilla-10000000-1-4]": 13.899,
"test_runner/performance/test_startup.py::test_startup": 890.114,
"test_runner/performance/test_startup.py::test_startup_simple": 2.51,
"test_runner/performance/test_wal_backpressure.py::test_heavy_write_workload[neon_off-10-5-5]": 527.245,
"test_runner/performance/test_wal_backpressure.py::test_heavy_write_workload[neon_on-10-5-5]": 583.46,
+49
View File
@@ -1,3 +1,6 @@
from contextlib import closing
import pytest
import requests
from fixtures.benchmark_fixture import MetricReport, NeonBenchmarker
from fixtures.neon_fixtures import NeonEnvBuilder
@@ -78,3 +81,49 @@ def test_startup_simple(neon_env_builder: NeonEnvBuilder, zenbenchmark: NeonBenc
# Imitate optimizations that console would do for the second start
endpoint.respec(skip_pg_catalog_updates=True)
# This test sometimes runs for longer than the global 5 minute timeout.
@pytest.mark.timeout(900)
def test_startup(neon_env_builder: NeonEnvBuilder, zenbenchmark: NeonBenchmarker):
neon_env_builder.num_safekeepers = 3
env = neon_env_builder.init_start()
# Start
env.neon_cli.create_branch("test_startup")
with zenbenchmark.record_duration("startup_time"):
endpoint = env.endpoints.create_start("test_startup")
endpoint.safe_psql("select 1;")
# Restart
endpoint.stop_and_destroy()
with zenbenchmark.record_duration("restart_time"):
endpoint.create_start("test_startup")
endpoint.safe_psql("select 1;")
# Fill up
num_rows = 1000000 # 30 MB
num_tables = 100
with closing(endpoint.connect()) as conn:
with conn.cursor() as cur:
for i in range(num_tables):
cur.execute(f"create table t_{i} (i integer);")
cur.execute(f"insert into t_{i} values (generate_series(1,{num_rows}));")
# Read
with zenbenchmark.record_duration("read_time"):
endpoint.safe_psql("select * from t_0;")
# Read again
with zenbenchmark.record_duration("second_read_time"):
endpoint.safe_psql("select * from t_0;")
# Restart
endpoint.stop_and_destroy()
with zenbenchmark.record_duration("restart_with_data"):
endpoint.create_start("test_startup")
endpoint.safe_psql("select 1;")
# Read
with zenbenchmark.record_duration("read_after_restart"):
endpoint.safe_psql("select * from t_0;")
+3 -3
View File
@@ -1,5 +1,5 @@
{
"postgres-v16": "825a713a4fc833a0924bf22ad34681de97c155a0",
"postgres-v15": "e45d092e4d2b09057671190a1c9d9292293cd6b7",
"postgres-v14": "1224330eee058ed6013840e2a6dace5af82150ac"
"postgres-v16": "550ffa6495a5dc62fccc3a8b449386633758680b",
"postgres-v15": "ab67ab96355d61e9d0218630be4aa7db53bf83e7",
"postgres-v14": "6669a672ee14ab2c09d44c4552f9a13fad3afc10"
}