## Problem
Vectored read path may return an image that's newer than the request lsn
under certain circumstances.
```
LSN
^
|
|
500 | ------------------------- -> branch point
400 | X
300 | X
200 | ------------------------------------> requested lsn
100 | X
|---------------------------------> Key
Legend:
* X - page images
```
The vectored read path inspects each ancestor timeline one by one
starting from the current one.
When moving into the ancestor timeline, the current code resets the
current search lsn (called `cont_lsn` in code)
to the lsn of the ancestor timeline
([here](d5708e7435/pageserver/src/tenant/timeline.rs (L2971))).
For instance, if the request lsn was 200, we would:
1. Look into the current timeline and find nothing for the key
2. Descend into the ancestor timeline and set `cont_lsn=500`
3. Return the page image at LSN 400
Myself and Christian find it very unlikely for this to have happened in
prod since the vectored read path
is always used at the last record lsn.
This issue was found by a regress test during the work to migrate get
page handling to use the vectored
implementation. I've applied my fix to that wip branch and it fixed the
issue.
## Summary of changes
The fix is to set the current search lsn to the min between the
requested LSN and the ancestor lsn.
Hence, at step 2 above we would set the current search lsn to 200 and
ignore the images above that.
A test illustrating the bug is also included. Fails without the patch
and passes with it.
#7030 introduced an annoying papercut, deeming a failure to acquire a
strong reference to `LayerInner` from `DownloadedLayer::drop` as a
canceled eviction. Most of the time, it wasn't that, but just timeline
deletion or tenant detach with the layer not wanting to be deleted or
evicted.
When a Layer is dropped as part of a normal shutdown, the `Layer` is
dropped first, and the `DownloadedLayer` the second. Because of this, we
cannot detect eviction being canceled from the `DownloadedLayer::drop`.
We can detect it from `LayerInner::drop`, which this PR adds.
Test case is added which before had 1 started eviction, 2 canceled. Now
it accurately finds 1 started, 1 canceled.
## Problem
Some tenants were observed to stop doing downloads after some time
## Summary of changes
- Fix a rogue `<` that was incorrectly scheduling work when `now` was
_before_ the scheduling target, rather than after. This usually resulted
in too-frequent execution, but could also result in never executing, if
the current time has advanced ahead of `next_download` at the time we
call `schedule()`.
- Fix in-memory list of timelines not being amended after timeline
deletion: the resulted in repeated harmless logs about the timeline
being removed, and redundant calls to remove_dir_all for the timeline
path.
- Add a log at startup to make it easier to see a particular tenant
starting in secondary mode (this is for parity with the logging that
exists when spawning an attached tenant). Previously searching on tenant
ID didn't provide a clear signal as to how the tenant was started during
pageserver start.
- Add a test that exercises secondary downloads using the background
scheduling, whereas existing tests were using the API hook to invoke
download directly.
For "timeline ancestor merge" or "timeline detach," we need to "cut"
delta layers at particular LSN. The name "truncate" is not used as it
would imply that a layer file changes, instead of what happens: we copy
keys with Lsn less than a "cut point".
Cc: #6994
Add the "copy delta layer prefix" operation to DeltaLayerInner, re-using
some of the vectored read internals. The code is `cfg(test)` until it
will be used later with a more complete integration test.
## Problem
This trace is emitted whenever a vectored read touches the end of a
delta layer file. It's a perfectly normal case, but I expected it to be
more rare when implementing the code.
## Summary of changes
Demote log to debug.
## Problem
We specify a bunch of possible error codes in the pageserver api swagger
spec. This is error prone and annoying to work with.
https://github.com/neondatabase/cloud/pull/11907 introduced generic
error handling on the control plane side, so we can now clean up the
spec.
## Summary of changes
* Remove generic error codes from swagger spec
* Update a couple route handlers which would previously return an error
without a `msg` field in the response body.
Tested via https://github.com/neondatabase/cloud/pull/12340
Related https://github.com/neondatabase/cloud/issues/7238
## Problem
test_sharding_smoke recently got an added section that checks deletion
of a sharded tenant. The storage controller does a retry loop for
deletion, waiting for a 404 response. When deletion is a bit slow (debug
builds), the retry of deletion was getting a 500 response -- this caused
the test to become flaky (example failure:
https://neon-github-public-dev.s3.amazonaws.com/reports/release-proxy/8659801445/index.html#testresult/b4cbf5b58190f60e/retries)
There was a false comment in the code:
```
match tenant.current_state() {
TenantState::Broken { .. } | TenantState::Stopping { .. } => {
- // If a tenant is broken or stopping, DeleteTenantFlow can
- // handle it: broken tenants proceed to delete, stopping tenants
- // are checked for deletion already in progress.
```
If the tenant is stopping, DeleteTenantFlow does not in fact handle it,
but returns a 500-yielding errror.
## Summary of changes
Before calling into DeleteTenantFlow, if the tenant is in
stopping|broken state then return 202 if a deletion is in progress. This
makes the API friendlier for retries.
The historic AlreadyInProgress (409) response still exists for if we
enter DeleteTenantFlow and unexpectedly see the tenant stopping. That
should go away when we implement #5080 . For the moment, callers that
handle 409s should continue to do so.
Before this PR, the `nix::poll::poll` call would stall the executor.
This PR refactors the `walredo::process` module to allow for different
implementations, and adds a new `async` implementation which uses
`tokio::process::ChildStd{in,out}` for IPC.
The `sync` variant remains the default for now; we'll do more testing in
staging and gradual rollout to prod using the config variable.
Performance
-----------
I updated `bench_walredo.rs`, demonstrating that a single `async`-based
walredo manager used by N=1...128 tokio tasks has lower latency and
higher throughput.
I further did manual less-micro-benchmarking in the real pageserver
binary.
Methodology & results are published here:
https://neondatabase.notion.site/2024-04-08-async-walredo-benchmarking-8c0ed3cc8d364a44937c4cb50b6d7019?pvs=4
tl;dr:
- use pagebench against a pageserver patched to answer getpage request &
small-enough working set to fit into PS PageCache / kernel page cache.
- compare knee in the latency/throughput curve
- N tenants, each 1 pagebench clients
- sync better throughput at N < 30, async better at higher N
- async generally noticable but not much worse p99.X tail latencies
- eyeballing CPU efficiency in htop, `async` seems significantly more
CPU efficient at ca N=[0.5*ncpus, 1.5*ncpus], worse than `sync` outside
of that band
Mental Model For Walredo & Scheduler Interactions
-------------------------------------------------
Walredo is CPU-/DRAM-only work.
This means that as soon as the Pageserver writes to the pipe, the
walredo process becomes runnable.
To the Linux kernel scheduler, the `$ncpus` executor threads and the
walredo process thread are just `struct task_struct`, and it will divide
CPU time fairly among them.
In `sync` mode, there are always `$ncpus` runnable `struct task_struct`
because the executor thread blocks while `walredo` runs, and the
executor thread becomes runnable when the `walredo` process is done
handling the request.
In `async` mode, the executor threads remain runnable unless there are
no more runnable tokio tasks, which is unlikely in a production
pageserver.
The above means that in `sync` mode, there is an implicit concurrency
limit on concurrent walredo requests (`$num_runtimes *
$num_executor_threads_per_runtime`).
And executor threads do not compete in the Linux kernel scheduler for
CPU time, due to the blocked-runnable-ping-pong.
In `async` mode, there is no concurrency limit, and the walredo tasks
compete with the executor threads for CPU time in the kernel scheduler.
If we're not CPU-bound, `async` has a pipelining and hence throughput
advantage over `sync` because one executor thread can continue
processing requests while a walredo request is in flight.
If we're CPU-bound, under a fair CPU scheduler, the *fixed* number of
executor threads has to share CPU time with the aggregate of walredo
processes.
It's trivial to reason about this in `sync` mode due to the
blocked-runnable-ping-pong.
In `async` mode, at 100% CPU, the system arrives at some (potentially
sub-optiomal) equilibrium where the executor threads get just enough CPU
time to fill up the remaining CPU time with runnable walredo process.
Why `async` mode Doesn't Limit Walredo Concurrency
--------------------------------------------------
To control that equilibrium in `async` mode, one may add a tokio
semaphore to limit the number of in-flight walredo requests.
However, the placement of such a semaphore is non-trivial because it
means that tasks queuing up behind it hold on to their request-scoped
allocations.
In the case of walredo, that might be the entire reconstruct data.
We don't limit the number of total inflight Timeline::get (we only
throttle admission).
So, that queue might lead to an OOM.
The alternative is to acquire the semaphore permit *before* collecting
reconstruct data.
However, what if we need to on-demand download?
A combination of semaphores might help: one for reconstruct data, one
for walredo.
The reconstruct data semaphore permit is dropped after acquiring the
walredo semaphore permit.
This scheme effectively enables both a limit on in-flight reconstruct
data and walredo concurrency.
However, sizing the amount of permits for the semaphores is tricky:
- Reconstruct data retrieval is a mix of disk IO and CPU work.
- If we need to do on-demand downloads, it's network IO + disk IO + CPU
work.
- At this time, we have no good data on how the wall clock time is
distributed.
It turns out that, in my benchmarking, the system worked fine without a
semaphore. So, we're shipping async walredo without one for now.
Future Work
-----------
We will do more testing of `async` mode and gradual rollout to prod
using the config flag.
Once that is done, we'll remove `sync` mode to avoid the temporary code
duplication introduced by this PR.
The flag will be removed.
The `wait()` for the child process to exit is still synchronous; the
comment [here](
655d3b6468/pageserver/src/walredo.rs (L294-L306))
is still a valid argument in favor of that.
The `sync` mode had another implicit advantage: from tokio's
perspective, the calling task was using up coop budget.
But with `async` mode, that's no longer the case -- to tokio, the writes
to the child process pipe look like IO.
We could/should inform tokio about the CPU time budget consumed by the
task to achieve fairness similar to `sync`.
However, the [runtime function for this is
`tokio_unstable`](`https://docs.rs/tokio/latest/tokio/task/fn.consume_budget.html).
Refs
----
refs #6628
refs https://github.com/neondatabase/neon/issues/2975
No functional changes, this is a comments/naming PR.
While merging sharding changes, some cleanup of the shard.rs types was
deferred.
In this PR:
- Rename `is_zero` to `is_shard_zero` to make clear that this method
doesn't literally mean that the entire object is zeros, just that it
refers to the 0th shard in a tenant.
- Pull definitions of types to the top of shard.rs and add a big comment
giving an overview of which type is for what.
Closes: https://github.com/neondatabase/neon/issues/6072
Adds another tool to the DR toolbox: ability in pagectl to
recover arbitrary prefixes in remote storage. Requires remote storage config,
the prefix, and the travel-to timestamp parameter
to be specified as cli args.
The done-if-after parameter is also supported.
Example invocation (after `aws login --profile dev`):
```
RUST_LOG=remote_storage=debug AWS_PROFILE=dev cargo run -p pagectl time-travel-remote-prefix 'remote_storage = { bucket_name = "neon-test-bucket-name", bucket_region = "us-east-2" }' wal/3aa8fcc61f6d357410b7de754b1d9001/641e5342083b2235ee3deb8066819683/ 2024-04-05T17:00:00Z
```
This has been written to resolve a customer recovery case:
https://neondb.slack.com/archives/C033RQ5SPDH/p1712256888468009
There is validation of the prefix to prevent accidentially specifying
too generic prefixes, which can cause corruption and data
loss if used wrongly. Still, the validation is not perfect and it is
important that the command is used with caution.
If possible, `time_travel_remote_storage` should
be used instead which has additional checks in place.
Problem
Currently, we base our time based layer rolling decision on the last
time we froze a layer. This means that if we roll a layer and then go
idle for longer than the checkpoint timeout the next layer will be
rolled after the first write. This is of course not desirable.
Summary of changes
Record the timepoint of the first write to an open layer and use that
for time based layer rolling decisions. Note that I had to keep
`Timeline::last_freeze_ts` for the sharded tenant disk consistent lsn
skip hack.
Fixes#7241
## Problem
We have two places that use a helper (`ser_rfc3339_millis`) to get serde
to stringify SystemTimes into the desired format.
## Summary of changes
Created a new module `utils::serde_system_time` and inside it a wrapper
type `SystemTime` for `std::time::SystemTime` that
serializes/deserializes to the RFC3339 format.
This new type is then used in the two places that were previously using
the helper for serialization, thereby eliminating the need to decorate
structs.
Closes#7151.
This PR is an off-by-default revision v2 of the (since-reverted) PR
#6555 / commit `3220f830b7fbb785d6db8a93775f46314f10a99b`.
See that PR for details on why running with a single runtime is
desirable and why we should be ready.
We reverted #6555 because it showed regressions in prodlike cloudbench,
see the revert commit message `ad072de4209193fd21314cf7f03f14df4fa55eb1`
for more context.
This PR makes it an opt-in choice via an env var.
The default is to use the 4 separate runtimes that we have today, there
shouldn't be any performance change.
I tested manually that the env var & added metric works.
```
# undefined env var => no change to before this PR, uses 4 runtimes
./target/debug/neon_local start
# defining the env var enables one-runtime mode, value defines that one runtime's configuration
NEON_PAGESERVER_USE_ONE_RUNTIME=current_thread ./target/debug/neon_local start
NEON_PAGESERVER_USE_ONE_RUNTIME=multi_thread:1 ./target/debug/neon_local start
NEON_PAGESERVER_USE_ONE_RUNTIME=multi_thread:2 ./target/debug/neon_local start
NEON_PAGESERVER_USE_ONE_RUNTIME=multi_thread:default ./target/debug/neon_local start
```
I want to use this change to do more manualy testing and potentially
testing in staging.
Future Work
-----------
Testing / deployment ergonomics would be better if this were a variable
in `pageserver.toml`.
It can be done, but, I don't need it right now, so let's stick with the
env var.
It's just unnecessary to use spawn_blocking there, and with
https://github.com/neondatabase/neon/pull/7331 , it will result in
really just one executor thread when enabling one-runtime with
current_thread executor.
We can currently underflow `pageserver_resident_physical_size_global`,
so the used disk bytes would show `u63::MAX` by mistake. The assumption
of the API (and the documented behavior) was to give the layer files
disk usage.
Switch to reporting numbers that match `df` output.
Fixes: #7336
## Problem
Ingest filtering wasn't being applied to timeline creations, so a
timeline created on a sharded tenant would use 20MB+ on each shard (each
shard got a full copy). This didn't break anything, but is inefficient
and leaves the system in a harder-to-validate state where shards
initially have some data that they will eventually drop during
compaction.
Closes: https://github.com/neondatabase/neon/issues/6649
## Summary of changes
- in `import_rel`, filter block-by-block with is_key_local
- During test_sharding_smoke, check that per-shard physical sizes are as
expected
- Also extend the test to check deletion works as expected (this was an
outstanding tech debt task)
## Problem
As https://github.com/neondatabase/neon/issues/6092 points out, this
test was (ab)using a failpoint!() with 'pause', which was occasionally
causing index uploads to get hung on a stuck executor thread, resulting
in timeouts waiting for remote_consistent_lsn.
That is one of several failure modes, but by far the most frequent.
## Summary of changes
- Replace the failpoint! with a `sleep_millis_async`, which is not only
async but also supports clean shutdown.
- Improve debugging: log the consistent LSN when scheduling an index
upload
- Tidy: remove an unnecessary checkpoint in the test code, where
last_flush_lsn_upload had just been called (this does a checkpoint
internally)
## Problem
The API client was written around the same time as some of the server
APIs changed from TenantId to TenantShardId
Closes: https://github.com/neondatabase/neon/issues/6154
## Summary of changes
- Refactor mgmt_api timeline_info and keyspace methods to use
TenantShardId to match the server
This doesn't make pagebench sharding aware, but it paves the way to do
so later.
## Problem
In the test for https://github.com/neondatabase/neon/pull/6776, a test
cases uses tiny layer sizes and tiny stripe sizes. This hits a scenario
where a shard's checkpoint interval spans a region where none of the
content in the WAL is ingested by this shard. Since there is no layer to
flush, we do not advance disk_consistent_lsn, and this causes the test
to fail while waiting for LSN to advance.
## Summary of changes
- Pass an LSN through `layer_flush_start_tx`. This is the LSN to which
we have frozen at the time we ask the flush to flush layers frozen up to
this point.
- In the layer flush task, if the layers we flush do not reach
`frozen_to_lsn`, then advance disk_consistent_lsn up to this point.
- In `maybe_freeze_ephemeral_layer`, handle the case where
last_record_lsn has advanced without writing a layer file: this ensures
that disk_consistent_lsn and remote_consistent_lsn advance anyway.
The net effect is that the disk_consistent_lsn is allowed to advance
past regions in the WAL where a shard ingests no data, and that we
uphold our guarantee that remote_consistent_lsn always eventually
reaches the tip of the WAL.
The case of no layer at all is hard to test at present due to >0 shards
being polluted with SLRU writes, but I have tested it locally with a
branch that disables SLRU writes on shards >0. We can tighten up the
testing on this in future as/when we refine shard filtering (currently
shards >0 need the SLRU because they use it to figure out cutoff in GC
using timestamp-to-lsn).
Some time ago, we had an issue where a deletion queue hang was also
causing timeline deletions to hang.
This was unnecessary because the timeline deletion doesn't _need_ to
flush the deletion queue, it just does it as a pleasantry to make the
behavior easier to understand and test.
In this PR, we wrap the flush calls in a 10 second timeout (typically
the flush takes milliseconds) so that in the event of issues with the
deletion queue, timeline deletions are slower but not entirely blocked.
Closes: https://github.com/neondatabase/neon/issues/6440
part of #6628
Before this PR, we used a std::sync::RwLock to coalesce multiple
callers on one walredo spawning. One thread would win the write lock
and others would queue up either at the read() or write() lock call.
In a scenario where a compute initiates multiple getpage requests
from different Postgres backends (= different page_service conns),
and we don't have a walredo process around, this means all these
page_service handler tasks will enter the spawning code path,
one of them will do the spawning, and the others will stall their
respective executor thread because they do a blocking
read()/write() lock call.
I don't know exactly how bad the impact is in reality because
posix_spawn uses CLONE_VFORK under the hood, which means that the
entire parent process stalls anyway until the child does `exec`,
which in turn resumes the parent.
But, anyway, we won't know until we fix this issue.
And, there's definitely a future way out of stalling the
pageserver on posix_spawn, namely, forking template walredo processes
that fork again when they need to be per-tenant.
This idea is tracked in
https://github.com/neondatabase/neon/issues/7320.
Changes
-------
This PR fixes that scenario by switching to use `heavier_once_cell`
for coalescing. There is a comment on the struct field that explains
it in a bit more nuance.
### Alternative Design
An alternative would be to use tokio::sync::RwLock.
I did this in the first commit in this PR branch,
before switching to `heavier_once_cell`.
Performance
-----------
I re-ran the `bench_walredo` and updated the results, showing that
the changes are neglible.
For the record, the earlier commit in this PR branch that uses
`tokio::sync::RwLock` also has updated benchmark numbers, and the
results / kinds of tiny regression were equivalent to
`heavier_once_cell`.
Note that the above doesn't measure performance on the cold path, i.e.,
when we need to launch the process and coalesce. We don't have a
benchmark
for that, and I don't expect any significant changes. We have metrics
and we log spawn latency, so, we can monitor it in staging & prod.
Risks
-----
As "usual", replacing a std::sync primitive with something that yields
to
the executor risks exposing concurrency that was previously implicitly
limited to the number of executor threads.
This would be the first one for walredo.
The risk is that we get descheduled while the reconstruct data is
already there.
That could pile up reconstruct data.
In practice, I think the risk is low because once we get scheduled
again, we'll
likely have a walredo process ready, and there is no further await point
until walredo is complete and the reconstruct data has been dropped.
This will change with async walredo PR #6548, and I'm well aware of it
in that PR.
This PR is a fallout from work on #7062.
# Changes
- Unify the freeze-and-flush and hard shutdown code paths into a single
method `Timeline::shutdown` that takes the shutdown mode as an argument.
- Replace `freeze_and_flush` bool arg in callers with that mode
argument, makes them more expressive.
- Switch timeline deletion to use `Timeline::shutdown` instead of its
own slightly-out-of-sync copy.
- Remove usage of `task_mgr::shutdown_watcher` /
`task_mgr::shutdown_token` where possible
# Future Work
Do we really need the freeze_and_flush?
If we could get rid of it, then there'd be no need for a specific
shutdown order.
Also, if you undo this patch's changes to the `eviction_task.rs` and
enable RUST_LOG=debug, it's easy to see that we do leave some task
hanging that logs under span `Connection{...}` at debug level. I think
it's a pre-existing issue; it's probably a broker client task.
## Problem
For reasons unrelated to this PR, I would like to make use of the tenant
conf in the `InMemoryLayer`. Previously, this was not possible without
copying and manually updating the copy to keep it in sync with updates.
## Summary of Changes:
Replace the `Arc<RwLock<AttachedTenantConf>>` with
`Arc<ArcSwap<AttachedTenantConf>>` (how many `Arc(s)` can one fit in a
type?). The most interesting part of this change is the updating of the
tenant config (`set_new_tenant_config` and
`set_new_location_config`). In theory, these two may race, although the
storage controller should prevent this via the tenant exclusive op lock.
Particular care has been taken to not "lose" a location config update by
using the read-copy-update approach when updating only the config.
## Problem
(Follows https://github.com/neondatabase/neon/pull/7237)
Some API users will query a tenant to wait for it to activate.
Currently, we return the current status of the tenant, whatever that may
be. Under heavy load, a pageserver starting up might take a long time to
activate such a tenant.
## Summary of changes
- In `tenant_status` handler, call wait_to_become_active on the tenant.
If the tenant is currently waiting for activation, this causes it to
skip the queue, similiar to other API handlers that require an active
tenant, like timeline creation. This avoids external services waiting a
long time for activation when polling GET /v1/tenant/<id>.
Tiered compaction hasn't scheduled the upload of image layers. In the
`test_gc_feedback.py` test this has caused warnings like with tiered
compaction:
```
INFO request[...] Deleting layer [...] not found in latest_files list, never uploaded?
```
Which caused errors like:
```
ERROR layer_delete[...] was unlinked but was not dangling
```
Fixes#7244
We want to move the code base away from task_mgr.
This PR refactors the walreceiver code such that it doesn't use
`task_mgr` anymore.
# Background
As a reminder, there are three tasks in a Timeline that's ingesting WAL.
`WalReceiverManager`, `WalReceiverConnectionHandler`, and
`WalReceiverConnectionPoller`.
See the documentation in `task_mgr.rs` for how they interact.
Before this PR, cancellation was requested through
task_mgr::shutdown_token() and `TaskHandle::shutdown`.
Wait-for-task-finish was implemented using a mixture of
`task_mgr::shutdown_tasks` and `TaskHandle::shutdown`.
This drawing might help:
<img width="300" alt="image"
src="https://github.com/neondatabase/neon/assets/956573/b6be7ad6-ecb3-41d0-b410-ec85cb8d6d20">
# Changes
For cancellation, the entire WalReceiver task tree now has a
`child_token()` of `Timeline::cancel`. The `TaskHandle` no longer is a
cancellation root.
This means that `Timeline::cancel.cancel()` is propagated.
For wait-for-task-finish, all three tasks in the task tree hold the
`Timeline::gate` open until they exit.
The downside of using the `Timeline::gate` is that we can no longer wait
for just the walreceiver to shut down, which is particularly relevant
for `Timeline::flush_and_shutdown`.
Effectively, it means that we might ingest more WAL while the
`freeze_and_flush()` call is ongoing.
Also, drive-by-fix the assertiosn around task kinds in `wait_lsn`. The
check for `WalReceiverConnectionHandler` was ineffective because that
never was a task_mgr task, but a TaskHandle task. Refine the assertion
to check whether we would wait, and only fail in that case.
# Alternatives
I contemplated (ab-)using the `Gate` by having a separate `Gate` for
`struct WalReceiver`.
All the child tasks would use _that_ gate instead of `Timeline::gate`.
And `struct WalReceiver` itself would hold an `Option<GateGuard>` of the
`Timeline::gate`.
Then we could have a `WalReceiver::stop` function that closes the
WalReceiver's gate, then drops the `WalReceiver::Option<GateGuard>`.
However, such design would mean sharing the WalReceiver's `Gate` in an
`Arc`, which seems awkward.
A proper abstraction would be to make gates hierarchical, analogous to
CancellationToken.
In the end, @jcsp and I talked it over and we determined that it's not
worth the effort at this time.
# Refs
part of #7062
The latest failures of test_secondary_downloads are spooky: layers are
missing on disk according to the test, but present according to the
pageserver logs:
- Make the pageserver assert that layers are really present on disk and
log the full path (debug mode only)
- Make the test dump a full listing on failure of the assert that failed
the last two times
Related: #6966
## Problem
The vectored read path holds the layer map lock while visiting a
timeline.
## Summary of changes
* Rework the fringe order to hold `Layer` on `Arc<InMemoryLayer>`
handles instead of descriptions that are resolved by the layer map at
the time of read. Note that previously `get_values_reconstruct_data` was
implemented for the layer description which already knew the lsn range
for the read. Now it is implemented on the new `ReadableLayer` handle
and needs to get the lsn range as an argument.
* Drop the layer map lock after updating the fringe.
Related https://github.com/neondatabase/neon/issues/6833
## Problem
Part of the legacy (but current) compaction algorithm is to find a stack
of overlapping delta layers which will be turned
into an image layer. This operation is exponential in terms of the
number of matching layers and we do it roughly every 20 seconds.
## Summary of changes
Only check if a new image layer is required if we've ingested a certain
amount of WAL since the last check.
The amount of wal is expressed in terms of multiples of checkpoint
distance, with the intuition being that
that there's little point doing the check if we only have two new L1
layers (not enough to create a new image).
## Problem
During this week's deployment we observed panics due to the blobs
for certain keys not fitting in the vectored read buffers. The likely
cause of this is a bloated AUX_FILE_KEY caused by logical replication.
## Summary of changes
This pr fixes the issue by allocating a buffer big enough to fit
the widest read. It also has the benefit of saving space if all keys
in the read have blobs smaller than the max vectored read size.
If the soft limit for the max size of a vectored read is violated,
we print a warning which includes the offending key and lsn.
A randomised (but deterministic) end to end test is also added for
vectored reads on the delta layer.
Many tests like `test_live_migration` or
`test_timeline_deletion_with_files_stuck_in_upload_queue` set
`compaction_threshold` to 1, to create a lot of changes/updates. The
compaction threshold was passed as `fanout` parameter to the
tiered_compaction function, which didn't support values of 1 however.
Now we change the assert to support it, while still retaining the
exponential nature of the increase in range in terms of lsn that a layer
is responsible for.
A large chunk of the failures in #6964 was due to hitting this issue
that we now resolved.
Part of #6768.
# Problem
As pointed out through doc-comments in this PR, `drop_old_connection` is
not cancellation-safe.
This means we can leave a `handle_walreceiver_connection` tokio task
dangling during Timeline shutdown.
More details described in the corresponding issue #7062.
# Solution
Don't cancel-by-drop the `connection_manager_loop_step` from the
`tokio::select!()` in the task_mgr task.
Instead, transform the code to use a `CancellationToken` ---
specifically, `task_mgr::shutdown_token()` --- and make code responsive
to it.
The `drop_old_connection()` is still not cancellation-safe and also
doesn't get a cancellation token, because there's no point inside the
function where we could return early if cancellation were requested
using a token.
We rely on the `handle_walreceiver_connection` to be sensitive to the
`TaskHandle`s cancellation token (argument name: `cancellation`).
Currently it checks for `cancellation` on each WAL message. It is
probably also sensitive to `Timeline::cancel` because ultimately all
that `handle_walreceiver_connection` does is interact with the
`Timeline`.
In summary, the above means that the following code (which is found in
`Timeline::shutdown`) now might **take longer**, but actually ensures
that all `handle_walreceiver_connection` tasks are finished:
```rust
task_mgr::shutdown_tasks(
Some(TaskKind::WalReceiverManager),
Some(self.tenant_shard_id),
Some(self.timeline_id)
)
```
# Refs
refs #7062
## Problem
This is a refactor.
This PR was a precursor to a much smaller change
e5bd602dc1,
where as I was writing it I found that we were not far from getting rid
of the last non-deprecated code paths that use `mgr::` scoped functions
to get at the TenantManager state.
We're almost done cleaning this up as per
https://github.com/neondatabase/neon/issues/5796. The only significant
remaining mgr:: item is `get_active_tenant_with_timeout`, which is
page_service's path for fetching tenants.
## Summary of changes
- Remove the bool argument to get_attached_tenant_shard: this was almost
always false from API use cases, and in cases when it was true, it was
readily replacable with an explicit check of the returned tenant's
status.
- Rather than letting the timeline eviction task query any tenant it
likes via `mgr::`, pass an `Arc<Tenant>` into the task. This is still an
ugly circular reference, but should eventually go away: either when we
switch to exclusively using disk usage eviction, or when we change
metadata storage to avoid the need to imitate layer accesses.
- Convert all the mgr::get_tenant call sites to use
TenantManager::get_attached_tenant_shard
- Move list_tenants into TenantManager.
## Problem
Follows: https://github.com/neondatabase/neon/pull/7182
- Sufficient concurrent writes could OOM a pageserver from the size of
indices on all the InMemoryLayer instances.
- Enforcement of checkpoint_period only happened if there were some
writes.
Closes: https://github.com/neondatabase/neon/issues/6916
## Summary of changes
- Add `ephemeral_bytes_per_memory_kb` config property. This controls the
ratio of ephemeral layer capacity to memory capacity. The weird unit is
to enable making the ratio less than 1:1 (set this property to 1024 to
use 1MB of ephemeral layers for every 1MB of RAM, set it smaller to get
a fraction).
- Implement background layer rolling checks in
Timeline::compaction_iteration -- this ensures we apply layer rolling
policy in the absence of writes.
- During background checks, if the total ephemeral layer size has
exceeded the limit, then roll layers whose size is greater than the mean
size of all ephemeral layers.
- Remove the tick() path from walreceiver: it isn't needed any more now
that we do equivalent checks from compaction_iteration.
- Add tests for the above.
---------
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
Currently, we return 409 (Conflict) in two cases:
- Temporary: Timeline creation cannot proceed because another timeline
with the same ID is being created
- Permanent: Timeline creation cannot proceed because another timeline
exists with different parameters but the same ID.
Callers which time out a request and retry should be able to distinguish
these cases.
Closes: #7208
## Summary of changes
- Expose `AlreadyCreating` errors as 429 instead of 409
## Problem
We currently hold the layer map read lock while doing IO on the read
path. This is not required for correctness.
## Summary of changes
Drop the layer map lock after figuring out which layer we wish to read
from.
Why is this correct:
* `Layer` models the lifecycle of an on disk layer. In the event the
layer is removed from local disk, it will be on demand downloaded
* `InMemoryLayer` holds the `EphemeralFile` which wraps the on disk
file. As long as the `InMemoryLayer` is in scope, it's safe to read from it.
Related https://github.com/neondatabase/neon/issues/6833
## Problem
Large quantities of ephemeral layer data can lead to excessive memory
consumption (https://github.com/neondatabase/neon/issues/6939). We
currently don't have a way to know how much ephemeral layer data is
present on a pageserver.
Before we can add new behaviors to proactively roll layers in response
to too much ephemeral data, we must calculate that total.
Related: https://github.com/neondatabase/neon/issues/6916
## Summary of changes
- Create GlobalResources and GlobalResourceUnits types, where timelines
carry a GlobalResourceUnits in their TimelineWriterState.
- Periodically update the size in GlobalResourceUnits:
- During tick()
- During layer roll
- During put() if the latest value has drifted more than 10MB since our
last update
- Expose the value of the global ephemeral layer bytes counter as a
prometheus metric.
- Extend the lifetime of TimelineWriterState:
- Instead of dropping it in TimelineWriter::drop, let it remain.
- Drop TimelineWriterState in roll_layer: this drops our guard on the
global byte count to reflect the fact that we're freezing the layer.
- Ensure the validity of the later in the writer state by clearing the
state in the same place we freeze layers, and asserting on the
write-ability of the layer in `writer()`
- Add a 'context' parameter to `get_open_layer_action` so that it can
skip the prev_lsn==lsn check when called in tick() -- this is needed
because now tick is called with a populated state, where
prev_lsn==Some(lsn) is true for an idle timeline.
- Extend layer rolling test to use this metric
Before this PR, each core had 3 executor threads from 3 different
runtimes. With this PR, we just have one runtime, with one thread per
core. Switching to a single tokio runtime should reduce that effective
over-commit of CPU and in theory help with tail latencies -- iff all
tokio tasks are well-behaved and yield to the runtime regularly.
Are All Tasks Well-Behaved? Are We Ready?
-----------------------------------------
Sadly there doesn't seem to be good out-of-the box tokio tooling to
answer this question.
We *believe* all tasks are well behaved in today's code base, as of the
switch to `virtual_file_io_engine = "tokio-epoll-uring"` in production
(https://github.com/neondatabase/aws/pull/1121).
The only remaining executor-thread-blocking code is walredo and some
filesystem namespace operations.
Filesystem namespace operations work is being tracked in #6663 and not
considered likely to actually block at this time.
Regarding walredo, it currently does a blocking `poll` for read/write to
the pipe file descriptors we use for IPC with the walredo process.
There is an ongoing experiment to make walredo async (#6628), but it
needs more time because there are surprisingly tricky trade-offs that
are articulated in that PR's description (which itself is still WIP).
What's relevant for *this* PR is that
1. walredo is always CPU-bound
2. production tail latencies for walredo request-response
(`pageserver_wal_redo_seconds_bucket`) are
- p90: with few exceptions, low hundreds of micro-seconds
- p95: except on very packed pageservers, below 1ms
- p99: all below 50ms, vast majority below 1ms
- p99.9: almost all around 50ms, rarely at >= 70ms
- [Dashboard
Link](https://neonprod.grafana.net/d/edgggcrmki3uof/2024-03-walredo-latency?orgId=1&var-ds=ZNX49CDVz&var-pXX_by_instance=0.9&var-pXX_by_instance=0.99&var-pXX_by_instance=0.95&var-adhoc=instance%7C%21%3D%7Cpageserver-30.us-west-2.aws.neon.tech&var-per_instance_pXX_max_seconds=0.0005&from=1711049688777&to=1711136088777)
The ones below 1ms are below our current threshold for when we start
thinking about yielding to the executor.
The tens of milliseconds stalls aren't great, but, not least because of
the implicit overcommit of CPU by the three runtimes, we can't be sure
whether these tens of milliseconds are inherently necessary to do the
walredo work or whether we could be faster if there was less contention
for CPU.
On the first item (walredo being always CPU-bound work): it means that
walredo processes will always compete with the executor threads.
We could yield, using async walredo, but then we hit the trade-offs
explained in that PR.
tl;dr: the risk of stalling executor threads through blocking walredo
seems low, and switching to one runtime cleans up one potential source
for higher-than-necessary stall times (explained in the previous
paragraphs).
Code Changes
------------
- Remove the 3 different runtime definitions.
- Add a new definition called `THE_RUNTIME`.
- Use it in all places that previously used one of the 3 removed
runtimes.
- Remove the argument from `task_mgr`.
- Fix failpoint usage where `pausable_failpoint!` should have been used.
We encountered some actual failures because of this, e.g., hung
`get_metric()` calls during test teardown that would client-timeout
after 300s.
As indicated by the comment above `THE_RUNTIME`, we could take this
clean-up further.
But before we create so much churn, let's first validate that there's no
perf regression.
Performance
-----------
We will test this in staging using the various nightly benchmark runs.
However, the worst-case impact of this change is likely compaction
(=>image layer creation) competing with compute requests.
Image layer creation work can't be easily generated & repeated quickly
by pagebench.
So, we'll simply watch getpage & basebackup tail latencies in staging.
Additionally, I have done manual benchmarking using pagebench.
Report:
https://neondatabase.notion.site/2024-03-23-oneruntime-change-benchmarking-22a399c411e24399a73311115fb703ec?pvs=4
Tail latencies and throughput are marginally better (no regression =
good).
Except in a workload with 128 clients against one tenant.
There, the p99.9 and p99.99 getpage latency is about 2x worse (at
slightly lower throughput).
A dip in throughput every 20s (compaction_period_ is clearly visible,
and probably responsible for that worse tail latency.
This has potential to improve with async walredo, and is an edge case
workload anyway.
Future Work
-----------
1. Once this change has shown satisfying results in production, change
the codebase to use the ambient runtime instead of explicitly
referencing `THE_RUNTIME`.
2. Have a mode where we run with a single-threaded runtime, so we
uncover executor stalls more quickly.
3. Switch or write our own failpoints library that is async-native:
https://github.com/neondatabase/neon/issues/7216
