## Problem
We don't have any benchmarks for Safekeeper WAL ingestion.
## Summary of changes
Add some basic benchmarks for WAL ingestion, specifically for
`SafeKeeper::process_msg()` (single append) and `WalAcceptor` (pipelined
batch ingestion). Also add some baseline file write benchmarks.
## Problem
When we use pull_timeline API on an evicted timeline, it gets downloaded
to serve the snapshot API request. That means that to evacuate all the
timelines from a node, the node needs enough disk space to download
partial segments from all timelines, which may not be physically the
case.
Closes: #8833
## Summary of changes
- Add a "try" variant of acquiring a residence guard, that returns None
if the timeline is offloaded
- During snapshot API handler, take a different code path if the
timeline isn't resident, where we just read the checkpoint and don't try
to read any segments.
## Problem
The storage components take an entire `SafekeeperConf` during
construction, but only actually use the `no_sync` field. This makes it
hard to understand the storage inputs (which fields do they actually
care about?), and is also inconvenient for tests and benchmarks that
need to set up a lot of unnecessary boilerplate.
## Summary of changes
* Don't take the entire config, but pass in the `no_sync` field
explicitly.
* Take the timeline dir instead of `ttid` as an input, since it's the
only thing it cares about.
* Fix a couple of tests to not leak tempdirs.
* Various minor tweaks.
Always do timeline init through atomic rename of temp directory. Add
GlobalTimelines::load_temp_timeline which does this, and use it from
both pull_timeline and basic timeline creation. Fixes a collection
of issues:
- previously timeline creation didn't really flushed cfile to disk
due to 'nothing to do if state didn't change' check;
- even if it did, without tmp dir it is possible to lose the cfile
but leave timeline dir in place, making it look corrupted;
- tenant directory creation fsync was missing in timeline creation;
- pull_timeline is now protected from concurrent both itself and
timeline creation;
- now global timelines map entry got special CreationInProgress
entry type which prevents from anyone getting access to timeline
while it is being created (previously one could get access to it,
but it was locked during creation, which is valid but confusing if
creation failed).
fixes#8927
When walproposer observes now higher term it restarts instead of
crashing whole compute with PANIC; this avoids compute crash after
term_bump call. After successfull election we're still checking
last_log_term of the highest given vote to ensure basebackup is good,
and PANIC otherwise.
It will be used for migration per
035-safekeeper-dynamic-membership-change.md
and
https://github.com/neondatabase/docs/pull/21
ref https://github.com/neondatabase/neon/issues/8700
Addresses the 1.82 beta clippy lint `too_long_first_doc_paragraph` by
adding newlines to the first sentence if it is short enough, and making
a short first sentence if there is the need.
Endpoint implementation sends msg to manager requesting to do the
reset. Manager stops current partial backup upload task if it exists and
performs the reset.
Also slightly tweak eviction condition: all full segments before
flush_lsn must be uploaded (and committed) and there must be only one
segment left on disk (partial). This allows to evict timelines which
started not on the first segment and didn't fill the whole
segment (previous condition wasn't good because last_removed_segno was
0).
ref https://github.com/neondatabase/neon/issues/8759
This commit tries to fix regular load spikes on staging, caused by too
many eviction and partial upload operations running at the same time.
Usually it was hapenning after restart, for partial backup the load was
delayed.
- Add a semaphore for evictions (2 permits by default)
- Rename `resident_since` to `evict_not_before` and smooth out the curve
by using random duration
- Use random duration in partial uploads as well
related to https://github.com/neondatabase/neon/issues/6338
some discussion in
https://neondb.slack.com/archives/C033RQ5SPDH/p1720601531744029
After https://github.com/neondatabase/neon/pull/8022 was deployed to
staging, I noticed many cases of timeouts. After inspecting the logs, I
realized that some operations are taking ~20 seconds and they're doing
while holding shared state lock. Usually it happens right after
redeploy, because compute reconnections put high load on disks. This
commit tries to improve observability around slow operations.
Non-observability changes:
- `TimelineState::finish_change` now skips update if nothing has changed
- `wal_residence_guard()` timeout is set to 30s
Fixes https://github.com/neondatabase/neon/issues/6337
Add safekeeper support to switch between `Present` and
`Offloaded(flush_lsn)` states. The offloading is disabled by default,
but can be controlled using new cmdline arguments:
```
--enable-offload
Enable automatic switching to offloaded state
--delete-offloaded-wal
Delete local WAL files after offloading. When disabled, they will be left on disk
--control-file-save-interval <CONTROL_FILE_SAVE_INTERVAL>
Pending updates to control file will be automatically saved after this interval [default: 300s]
```
Manager watches state updates and detects when there are no actvity on
the timeline and actual partial backup upload in remote storage. When
all conditions are met, the state can be switched to offloaded.
In `timeline.rs` there is `StateSK` enum to support switching between
states. When offloaded, code can access only control file structure and
cannot use `SafeKeeper` to accept new WAL.
`FullAccessTimeline` is now renamed to `WalResidentTimeline`. This
struct contains guard to notify manager about active tasks requiring
on-disk WAL access. All guards are issued by the manager, all requests
are sent via channel using `ManagerCtl`. When manager receives request
to issue a guard, it unevicts timeline if it's currently evicted.
Fixed a bug in partial WAL backup, it used `term` instead of
`last_log_term` previously.
After this commit is merged, next step is to roll this change out, as in
issue #6338.
- Add /snapshot http endpoing streaming tar archive timeline contents up to
flush_lsn.
- Add check that term doesn't change, corresponding test passes now.
- Also prepares infra to hold off WAL removal during the basebackup.
- Sprinkle fsyncs to persist the pull_timeline result.
ref https://github.com/neondatabase/neon/issues/6340
This is a preparation for
https://github.com/neondatabase/neon/issues/6337.
The idea is to add FullAccessTimeline, which will act as a guard for
tasks requiring access to WAL files. Eviction will be blocked on these
tasks and WAL won't be deleted from disk until there is at least one
active FullAccessTimeline.
To get FullAccessTimeline, tasks call `tli.full_access_guard().await?`.
After eviction is implemented, this function will be responsible for
downloading missing WAL file and waiting until the download finishes.
This commit also contains other small refactorings:
- Separate `get_tenant_dir` and `get_timeline_dir` functions for
building a local path. This is useful for looking at usages and finding
tasks requiring access to local filesystem.
- `timeline_manager` is now responsible for spawning all background
tasks
- WAL removal task is now spawned instantly after horizon is updated
epoch is a historical and potentially confusing name. It semantically means
lastLogTerm from the raft paper, so let's use it.
This commit changes only internal namings, not public interface (http).
Do pull_timeline while WAL is being removed. To this end
- extract pausable_failpoint to utils, sprinkle pull_timeline with it
- add 'checkpoint' sk http endpoint to force WAL removal.
After fixing checking for pull file status code test fails so far which is
expected.
I looked at the metrics from
https://github.com/neondatabase/neon/pull/7768 on staging and it seems
that manager does too many iterations. This is probably caused by
background job `remove_wal.rs` which iterates over all timelines and
tries to remove WAL and persist control file. This causes shared state
updates and wakes up the manager. The fix is to skip notifying about the
updates if nothing was updated.
## Problem
Safekeeper Timeline uses a channel for cancellation, but we have a
dedicated type for that.
## Summary of changes
- Use CancellationToken in Timeline
In safekeepers we have several background tasks. Previously `WAL backup`
task was spawned by another task called `wal_backup_launcher`. That task
received notifications via `wal_backup_launcher_rx` and decided to spawn
or kill existing backup task associated with the timeline. This was
inconvenient because each code segment that touched shared state was
responsible for pushing notification into `wal_backup_launcher_tx`
channel. This was error prone because it's easy to miss and could lead
to deadlock in some cases, if notification pushing was done in the wrong
order.
We also had a similar issue with `is_active` timeline flag. That flag
was calculated based on the state and code modifying the state had to
call function to update the flag. We had a few bugs related to that,
when we forgot to update `is_active` flag in some places where it could
change.
To fix these issues, this PR adds a new `timeline_manager` background
task associated with each timeline. This task is responsible for
managing all background tasks, including `is_active` flag which is used
for pushing broker messages. It is subscribed for updates in timeline
state in a loop and decides to spawn/kill background tasks when needed.
There is a new structure called `TimelinesSet`. It stores a set of
`Arc<Timeline>` and allows to copy the set to iterate without holding
the mutex. This is what replaced `is_active` flag for the broker. Now
broker push task holds a reference to the `TimelinesSet` with active
timelines and use it instead of iterating over all timelines and
filtering by `is_active` flag.
Also added some metrics for manager iterations and active backup tasks.
Ideally manager should be doing not too many iterations and we should
not have a lot of backup tasks spawned at the same time.
Fixes#7751
---------
Co-authored-by: Arseny Sher <sher-ars@yandex.ru>
To avoid pageserver gc'ing data needed by standby, propagate standby apply LSN
through standby -> safekeeper -> broker -> pageserver flow and hold off GC for
it. Iteration of GC resets the value to remove the horizon when standby goes
away -- pushes are assumed to happen at least once between gc iterations. As a
safety guard max allowed lag compared to normal GC horizon is hardcoded as 10GB.
Add test for the feature.
Co-authored-by: Konstantin Knizhnik <knizhnik@neon.tech>
Add support for backing up partial segments to remote storage. Disabled
by default, can be enabled with `--partial-backup-enabled`.
Safekeeper timeline has a background task which is subscribed to
`commit_lsn` and `flush_lsn` updates. After the partial segment was
updated (`flush_lsn` was changed), the segment will be uploaded to S3 in
about 15 minutes.
The filename format for partial segments is
`Segment_Term_Flush_Commit_skNN.partial`, where:
- `Segment` – the segment name, like `000000010000000000000001`
- `Term` – current term
- `Flush` – flush_lsn in hex format `{:016X}`, e.g. `00000000346BC568`
- `Commit` – commit_lsn in the same hex format
- `NN` – safekeeper_id, like `1`
The full object name example:
`000000010000000000000002_2_0000000002534868_0000000002534410_sk1.partial`
Each safekeeper will keep info about remote partial segments in its
control file. Code updates state in the control file before doing any S3
operations. This way control file stores information about all
potentially existing remote partial segments and can clean them up after
uploading a newer version.
Closes#6336
Previously we aggregated ps_feedback on each safekeeper and sent it to
walproposer with every AppendResponse. This PR changes it to send
ps_feedback to walproposer right after receiving it from pageserver,
without aggregating it in memory. Also contains some preparations for
implementing backpressure support for sharding.
Add `--walsenders-keep-horizon` argument to safekeeper cmdline. It will
prevent deleting WAL segments from disk if they are needed by the active
START_REPLICATION connection.
This is useful for sharding. Without this option, if one of the shard
falls behind, it starts to read WAL from S3, which is much slower than
disk. This can result in huge shard lagging.
In the most straightforward way; safekeeper performs it in DELETE endpoint
implementation, with no coordination between sks.
delete_force endpoint in the code is renamed to delete as there is only one way
to delete.
safekeeper.rs is mostly about consensus, but state is wider. Also form
SafekeeperState which encapsulates persistent part + in memory layer with API
for atomic updates.
Moves remote_consistent_lsn back to SafekeeperMemState, fixes its absense from
memory dump.
Also renames SafekeeperState to TimelinePersistentState, as TimelineMemState and
TimelinePersistent state are created.
Implement API for cloning a single timeline inside a safekeeper. Also
add API for calculating a sha256 hash of WAL, which is used in tests.
`/copy` API works by copying objects inside S3 for all but the last
segments, and the last segments are copied on-disk. A special temporary
directory is created for a timeline, because copy can take a lot of
time, especially for large timelines. After all files segments have been
prepared, this directory is mounted to the main tree and timeline is
loaded to memory.
Some caveats:
- large timelines can take a lot of time to copy, because we need to
copy many S3 segments
- caller should wait for HTTP call to finish indefinetely and don't
close the HTTP connection, because it will stop the process, which is
not continued in the background
- `until_lsn` must be a valid LSN, otherwise bad things can happen
- API will return 200 if specified `timeline_id` already exists, even if
it's not a copy
- each safekeeper will try to copy S3 segments, so it's better to not
call this API in-parallel on different safekeepers
Improve the serde impl for several types (`Lsn`, `TenantId`,
`TimelineId`) by making them sensitive to
`Serializer::is_human_readadable` (true for json, false for bincode).
Fixes#3511 by:
- Implement the custom serde for `Lsn`
- Implement the custom serde for `Id`
- Add the helper module `serde_as_u64` in `libs/utils/src/lsn.rs`
- Remove the unnecessary attr `#[serde_as(as = "DisplayFromStr")]` in
all possible structs
Additionally some safekeeper types gained serde tests.
---------
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
Implements fetching of WAL by safekeeper from another safekeeper by imitating
behaviour of last elected leader. This allows to avoid WAL accumulation on
compute and facilitates faster compute startup as it doesn't need to download
any WAL. Actually removing WAL download in walproposer is a matter of another
patch though.
There is a per timeline task which always runs, checking regularly if it should
start recovery frome someone, meaning there is something to fetch and there is
no streaming compute. It then proceeds with fetching, finishing when there is
nothing more to receive.
Implements https://github.com/neondatabase/neon/pull/4875
Doesn't do an upgrade of rustc to 1.73.0 as we want to wait for the
cargo response of the curl CVE before updating. In preparation for an
update, we address the clippy lints that are newly firing in 1.73.0.
Fixes#4689 by replacing all of `std::Path` , `std::PathBuf` with
`camino::Utf8Path`, `camino::Utf8PathBuf` in
- pageserver
- safekeeper
- control_plane
- libs/remote_storage
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
Instead of fixed during the start of replication. To this end, create
term_flush_lsn watch channel similar to commit_lsn one. This allows to continue
recovery streaming if new data appears.
Slightly refactors init: now load_tenant_timelines is also async to properly
init the timeline, but to keep global map lock sync we just acquire it anew for
each timeline.
Recovery task itself is just a stub here.
part of
https://github.com/neondatabase/neon/pull/4875
It allows term leader to ensure he pulls data from the correct term. Absense of
it wasn't very problematic due to CRC checks, but let's be strict.
walproposer still doesn't use it as we're going to remove recovery completely
from it.
## Problem
The safekeeper advertises the same address specified in `--listen-pg`,
which is problematic when the listening address is different from the
address that the pageserver can use to connect to the safekeeper.
## Summary of changes
Add a new optional flag called `--advertise-pg` for the address to be
advertised. If this flag is not specified, the behavior is the same as
before.
## Problem
During timeline creation we create special mark file which presense
indicates that initialization didnt complete successfully. In case of a
crash restart we can remove such half-initialized timeline and following
retry from control plane side should perform another attempt.
So in case of a possible crash restart during initial loading we have
following picture:
```
timelines
| - <timeline_id>___uninit
| - <timeline_id>
| - | <timeline files>
```
We call `std::fs::read_dir` to walk files in `timelines` directory one
by one. If we see uninit file
we proceed with deletion of both, timeline directory and uninit file. If
we see timeline we check if uninit file exists and do the same cleanup.
But in fact its possible to get both branches to be true at the same
time. Result of readdir doesnt reflect following directory state
modifications. So you can still get "valid" entry on the next iteration
of the loop despite the fact that it was deleted in one of the previous
iterations of the loop.
To see that you can apply the following patch (it disables uninit mark
cleanup on successful timeline creation):
```diff
diff --git a/pageserver/src/tenant.rs b/pageserver/src/tenant.rs
index 4beb2664..b3cdad8f 100644
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
@@ -224,11 +224,6 @@ impl UninitializedTimeline<'_> {
)
})?;
}
- uninit_mark.remove_uninit_mark().with_context(|| {
- format!(
- "Failed to remove uninit mark file for timeline {tenant_id}/{timeline_id}"
- )
- })?;
v.insert(Arc::clone(&new_timeline));
new_timeline.maybe_spawn_flush_loop();
```
And perform the following steps:
```bash
neon_local init
neon_local start
neon_local tenant create
neon_local stop
neon_local start
```
The error is:
```log
INFO load{tenant_id=X}:blocking: Found an uninit mark file .neon/tenants/X/timelines/Y.___uninit, removing the timeline and its uninit mark
2023-06-09T18:43:41.664247Z ERROR load{tenant_id=X}: load failed, setting tenant state to Broken: failed to load metadata
Caused by:
0: Failed to read metadata bytes from path .neon/tenants/X/timelines/Y/metadata
1: No such file or directory (os error 2)
```
So uninit mark got deleted together with timeline directory but we still
got directory entry for it and tried to load it.
The bug prevented tenant from being successfully loaded.
## Summary of changes
Ideally I think we shouldnt place uninit marks in the same directory as timeline directories but move them to separate directory and
gather them as an input to actual listing, but that would be sort of an
on-disk format change, so just check whether entries are still valid
before operating on them.
This is a full switch, fs io operations are also tokio ones, working through
thread pool. Similar to pageserver, we have multiple runtimes for easier `top`
usage and isolation.
Notable points:
- Now that guts of safekeeper.rs are full of .await's, we need to be very
careful not to drop task at random point, leaving timeline in unclear
state. Currently the only writer is walreceiver and we don't have top
level cancellation there, so we are good. But to be safe probably we should
add a fuse panicking if task is being dropped while operation on a timeline
is in progress.
- Timeline lock is Tokio one now, as we do disk IO under it.
- Collecting metrics got a crutch: since prometheus Collector is
synchronous, it spawns a thread with current thread runtime collecting data.
- Anything involving closures becomes significantly more complicated, as
async fns are already kinda closures + 'async closures are unstable'.
- Main thread now tracks other main tasks, which got much easier.
- The only sync place left is initial data loading, as otherwise clippy
complains on timeline map lock being held across await points -- which is
not bad here as it happens only in single threaded runtime of main thread.
But having it sync doesn't hurt either.
I'm concerned about performance of thread pool io offloading, async traits and
many await points; but we can try and see how it goes.
fixes https://github.com/neondatabase/neon/issues/3036
fixes https://github.com/neondatabase/neon/issues/3966
It should make remote_consistent_lsn commonly up-to-date on non actively writing
projects, which removes spike or pageserver -> safekeeper reconnections on
storage nodes restart.
Add HTTP endpoint to initialize safekeeper timeline from peer
safekeepers. This is useful for initializing new safekeeper to replace
failed safekeeper. Not fully "correct" in all cases, but should work in
most.
This code is not suitable for production workloads but can be tested on
staging to get started. New endpoint is separated from usual cases and
should not affect anything if no one explicitly uses a new endpoint. We
can rollback this commit in case of issues.
Refactors walsenders out of timeline.rs to makes it less convoluted into
separate WalSenders with its own lock, but otherwise having the same structure.
Tracking of in-memory remote_consistent_lsn is also moved there as it is mainly
received from pageserver.
State of walsender (feedback) is also restructured to be cleaner; now it is
either PageserverFeedback or StandbyFeedback(StandbyReply, HotStandbyFeedback),
but not both.
Otherwise it can lag a lot, preventing WAL segments cleanup. Also max
wal_backup_lsn on update, pulling it down is pointless.
Should help with https://github.com/neondatabase/neon/issues/3957, but
will not fix it completely.
