## Problem
Previously, `GET /v1/tenant/:tenant_id/timeline` and `GET
/v1/tenant/:tenant_id/timeline/:timeline_id`
would bump the priority of the background task which computes the
initial logical size by cancelling
the wait on the synchronisation semaphore. However, the request would
still return an approximate
logical size. It's undesirable to force background work for a status
request.
## Summary of changes
This PR updates the priority used by the timeline status request such
that they don't do priority boosting
by default anymore. An optional query parameter,
`force-await-initial-logical-size`, is added for both
mentioned endpoints. When set to true, it will skip the concurrency
limiting semaphore and wait
for the background task to complete before returning the exact logical
size.
In order to exercise this behaviour in a test I had to add an extra
failpoint. If you think it's too intrusive,
it can be removed.
Also fixeda small bug where the cancellation of a download is reported as an
opaque download failure upstream. This caused `test_location_conf_churn`
to fail at teardown due to a WARN log line.
Closes https://github.com/neondatabase/neon/issues/6168
Random find while looking at an idle 20k tenant pageserver where each
tenant
has 9 tiny L0 layers and compaction produces no new L1s / image layers.
The aggregate CPU cost of running this every 20s for 20k tenants is
actually substantial, due to the use of `spawn_blocking`.
This reverts commit ab1f37e908.
Thereby
fixes#5479
Updated Analysis
================
The problem with the original patch was that it, for the first time,
exposed the `VirtualFile` code to tokio task concurrency instead of just
thread-based concurrency. That caused the VirtualFile file descriptor
cache to start thrashing, effectively grinding the system to a halt.
Details
-------
At the time of the original patch, we had a _lot_ of runnable tasks in
the pageserver.
The symptom that prompted the revert (now being reverted in this PR) is
that our production systems fell into a valley of zero goodput, high
CPU, and zero disk IOPS shortly after PS restart.
We lay out the root cause for that behavior in this subsection.
At the time, there was no concurrency limit on the number of concurrent
initial logical size calculations.
Initial size calculation was initiated for all timelines within the
first 10 minutes as part of consumption metrics collection.
On a PS with 20k timelines, we'd thus have 20k runnable tasks.
Before the original patch, the `VirtualFile` code never returned
`Poll::Pending`.
That meant that once we entered it, the calling tokio task would not
yield to the tokio executor until we were done performing the
VirtualFile operation, i.e., doing a blocking IO system call.
The original patch switched the VirtualFile file descriptor cache's
synchronization primitives to those from `tokio::sync`.
It did not change that we were doing synchronous IO system calls.
And the cache had more slots than we have tokio executor threads.
So, these primitives never actually needed to return `Poll::Pending`.
But, the tokio scheduler makes tokio sync primitives return `Pending`
*artificially*, as a mechanism for the scheduler to get back into
control more often
([example](https://docs.rs/tokio/1.35.1/src/tokio/sync/batch_semaphore.rs.html#570)).
So, the new reality was that VirtualFile calls could now yield to the
tokio executor.
Tokio would pick one of the other 19999 runnable tasks to run.
These tasks were also using VirtualFile.
So, we now had a lot more concurrency in that area of the code.
The problem with more concurrency was that caches started thrashing,
most notably the VirtualFile file descriptor cache: each time a task
would be rescheduled, it would want to do its next VirtualFile
operation. For that, it would first need to evict another (task's)
VirtualFile fd from the cache to make room for its own fd. It would then
do one VirtualFile operation before hitting an await point and yielding
to the executor again. The executor would run the other 19999 tasks for
fairness before circling back to the first task, which would find its fd
evicted.
The other cache that would theoretically be impacted in a similar way is
the pageserver's `PageCache`.
However, for initial logical size calculation, it seems much less
relevant in experiments, likely because of the random access nature of
initial logical size calculation.
Fixes
=====
We fixed the above problems by
- raising VirtualFile cache sizes
- https://github.com/neondatabase/cloud/issues/8351
- changing code to ensure forward-progress once cache slots have been
acquired
- https://github.com/neondatabase/neon/pull/5480
- https://github.com/neondatabase/neon/pull/5482
- tbd: https://github.com/neondatabase/neon/issues/6065
- reducing the amount of runnable tokio tasks
- https://github.com/neondatabase/neon/pull/5578
- https://github.com/neondatabase/neon/pull/6000
- fix bugs that caused unnecessary concurrency induced by connection
handlers
- https://github.com/neondatabase/neon/issues/5993
I manually verified that this PR doesn't negatively affect startup
performance as follows:
create a pageserver in production configuration, with 20k
tenants/timelines, 9 tiny L0 layer files each; Start it, and observe
```
INFO Startup complete (368.009s since start) elapsed_ms=368009
```
I further verified in that same setup that, when using `pagebench`'s
getpage benchmark at as-fast-as-possible request rate against 5k of the
20k tenants, the achieved throughput is identical. The VirtualFile cache
isn't thrashing in that case.
Future Work
===========
We will still exposed to the cache thrashing risk from outside factors,
e.g., request concurrency is unbounded, and initial size calculation
skips the concurrency limiter when we establish a walreceiver
connection.
Once we start thrashing, we will degrade non-gracefully, i.e., encounter
a valley as was seen with the original patch.
However, we have sufficient means to deal with that unlikely situation:
1. we have dashboards & metrics to monitor & alert on cache thrashing
2. we can react by scaling the bottleneck resources (cache size) or by
manually shedding load through tenant relocation
Potential systematic solutions are future work:
* global concurrency limiting
* per-tenant rate limiting => #5899
* pageserver-initiated load shedding
Related Issues
==============
This PR unblocks the introduction of tokio-epoll-uring for asynchronous
disk IO ([Epic](#4744)).
PR #6266 broke the getpage_latest_lsn benchmark.
Before this patch, we'd fail with
```
not implemented: split up range
```
because `r.start = rel size key` and `r.end = rel size key + 1`.
The filtering of the key ranges in that loop is a bit ugly, but,
I measured:
* setup with 180k layer files (20k tenants * 9 layers).
* total physical size is 463GiB
* 5k tenants, the range filtering takes `0.6 seconds` on an
i3en.3xlarge.
That's a tiny fraction of the overall time it takes for pagebench to get
ready to send requests. So, this is good enough for now / there are
other bottlenecks that are bigger.
## Problem
The code for tenant create and tenant attach was just a special case of
what upsert_location does.
## Summary of changes
- Use `upsert_location` for create and attach APIs
- Clean up error handling in upsert_location so that it can generate
appropriate HTTP response codes
- Update tests that asserted the old non-idempotent behavior of attach
- Rework the `test_ignore_while_attaching` test, and fix tenant shutdown
during activation, which this test was supposed to cover, but it was
actually just waiting for activation to complete.
This uses the [newly stable](https://blog.rust-lang.org/2023/12/21/async-fn-rpit-in-traits.html)
async trait feature for three internal traits. One requires `Send`
bounds to be present so uses `impl Future<...> + Send` instead.
Advantages:
* less macro usage
* no extra boxing
Disadvantages:
* impl syntax needed for `Send` bounds is a bit more verbose (but only
required in one place)
This fixes the clippy lint firing on macOS on the conversion which
needed for portability. For some reason, the logic in
https://github.com/rust-lang/rust-clippy/pull/11669 to avoid an overlap
is not working.
These functions don't need updating for sharding: it's fine for them to
remain shard-naive, as they're only used in the context of dumping a
layer file. The sharding metadata doesn't live in the layer file, it
lives in the index.
## Problem
Noticed while making other changes that there were `pub` items that were
unused.
## Summary of changes
- Make everything `pub(crate)` in metrics.rs, apart from items used from
`bin/`
- Fix the timelines eviction metric: it was never being incremented
- Remove an unused ephemeral_bytes counter.
## Problem
- When a client requests a key that isn't found in any shard on the node
(edge case that only happens if a compute's config is out of date), we
should prompt them to reconnect (as this includes a backoff), since they
will not be able to complete the request until they eventually get a
correct pageserver connection string.
- QueryError::Other is used excessively: this contains a type-ambiguous
anyhow::Error and is logged very verbosely (including backtrace).
## Summary of changes
- Introduce PageStreamError to replace use of anyhow::Error in request
handlers for getpage, etc.
- Introduce Reconnect and NotFound variants to QueryError
- Map the "shard routing error" case to PageStreamError::Reconnect ->
QueryError::Reconnect
- Update type conversions for LSN timeouts and tenant/timeline not found
errors to use PageStreamError::NotFound->QueryError::NotFound
Previously, we would wait for the LSN to be visible on whichever
timeline we happened to load at the start of the connection, then
proceed to look up the correct timeline for the key and do the read.
If the timeline holding the key was behind the timeline we used
for the LSN wait, then we might serve an apparently-successful read result
that actually contains data from behind the requested lsn.
If there is some secondary shard for a tenant on the same
node as an attached shard, the secondary shard could trip up
this code and cause page_service to incorrectly
get an error instead of finding the attached shard.
## Problem
For context, this problem was observed in a research project where we
try to make neon run in multiple regions and I was asked by @hlinnaka to
make this PR.
In our project, we use the pageserver in a non-conventional way such
that we would send a larger number of requests to the pageserver than
normal (imagine postgres without the buffer pool). I measured the time
from the moment a WAL record left the safekeeper to when it reached the
pageserver
([code](e593db1f5a/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs (L282-L287)))
and observed that when the number of get_page_at_lsn requests was high,
the wal receiving time increased significantly (see the left side of the
graphs below).
Upon further investigation, I found that the delay was caused by this
line
d2ca410919/pageserver/src/tenant/timeline.rs (L2348)
The `get_layer_for_write` method is called for every value during WAL
ingestion and it tries to acquire layers write lock every time, thus
this results in high contention when read lock is acquired more
frequently.
## Summary of changes
It is unnecessary to call `get_layer_for_write` repeatedly for all
values in a WAL message since they would end up in the same memory layer
anyway, so I created the batched versions of `InMemoryLayer::put_value`,
`InMemoryLayer ::put_tombstone`, `Timeline::put_value`, and
`Timeline::put_tombstone`, that acquire the locks once for a batch of
values.
Additionally, `DatadirModification` is changed to store multiple
versions of uncommitted values, and `WalIngest::ingest_record()` can now
ingest records without immediately committing them.
With these new APIs, the new ingestion loop can be changed to commit for
every `ingest_batch_size` records. The `ingest_batch_size` variable is
exposed as a config. If it is set to 1 then we get the same behavior
before this change. I found that setting this value to 100 seems to work
the best, and you can see its effect on the right side of the above
graphs.
---------
Co-authored-by: John Spray <john@neon.tech>
## Problem
During startup, a tenant delete request might have to retry for many
minutes waiting for a tenant to enter Active state.
## Summary of changes
- Refactor delete_tenant into TenantManager: this is not a functional
change, but will avoid merge conflicts with
https://github.com/neondatabase/neon/pull/6105 later
- Add 412 responses to the swagger definition of this endpoint.
- Use Tenant::wait_to_become_active in `TenantManager::delete_tenant`
---------
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
This PR adds a component-level benchmarking utility for pageserver.
Its name is `pagebench`.
The problem solved by `pagebench` is that we want to put Pageserver
under high load.
This isn't easily achieved with `pgbench` because it needs to go through
a compute, which has signficant performance overhead compared to
accessing Pageserver directly.
Further, compute has its own performance optimizations (most
importantly: caches). Instead of designing a compute-facing workload
that defeats those internal optimizations, `pagebench` simply bypasses
them by accessing pageserver directly.
Supported benchmarks:
* getpage@latest_lsn
* basebackup
* triggering logical size calculation
This code has no automated users yet.
A performance regression test for getpage@latest_lsn will be added in a
later PR.
part of https://github.com/neondatabase/neon/issues/5771
Adds a new disk usage based eviction option, EvictionOrder, which
selects whether to use the current `AbsoluteAccessed` or this new
proposed but not yet tested `RelativeAccessed`. Additionally a fudge
factor was noticed while implementing this, which might help sparing
smaller tenants at the expense of targeting larger tenants.
Cc: #5304
Co-authored-by: Arpad Müller <arpad@neon.tech>
This is a precursor to:
- https://github.com/neondatabase/neon/pull/6185
While that PR contains big changes to neon_local and attachment_service,
this PR contains a few unrelated standalone changes generated while
working on that branch:
- Fix restarting a pageserver when it contains multiple shards for the
same tenant
- When using location_config api to attach a tenant, create its
timelines dir
- Update test paths where generations were previously optional to make
them always-on: this avoids tests having to spuriously assert that
attachment_service is not None in order to make the linter happy.
- Add a TenantShardId python implementation for subsequent use in test
helpers that will be made shard-aware
- Teach scrubber to read across shards when checking for layer
existence: this is a refactor to track the list of existent layers at
tenant-level rather than locally to each timeline. This is a precursor
to testing shard splitting.
It wasn't being incremented.
Fixup of
commit 1c88824ed0
Author: Christian Schwarz <christian@neon.tech>
Date: Fri Dec 1 12:52:59 2023 +0100
initial logical size calculation: add a bunch of metrics (#5995)
This squashes a particularly noisy warn-level log that occurs when
safekeepers are restarted.
Unfortunately the error type from `tonic` doesn't provide a neat way of
matching this, so we use a string comparison
Part of getpage@lsn benchmark epic:
https://github.com/neondatabase/neon/issues/5771
This allows getting the list of tenants and timelines without triggering
initial logical size calculation by requesting the timeline details API
response, which would skew our results.
## Problem
#6112 added some logs and metrics: clean these up a bit:
- Avoid counting startup completions for tenants launched after startup
- exclude no-op cases from timing histograms
- remove a rogue log messages
## Problem
During startup, a client request might have to wait a long time while
the system is busy initializing all the attached tenants, even though
most of the attached tenants probably don't have any client requests to
service, and could wait a bit.
## Summary of changes
- Add a semaphore to limit how many Tenant::spawn()s may concurrently do
I/O to attach their tenant (i.e. read indices from remote storage, scan
local layer files, etc).
- Add Tenant::activate_now, a hook for kicking a tenant in its spawn()
method to skip waiting for the warmup semaphore
- For tenants that attached via warmup semaphore units, wait for logical
size calculation to complete before dropping the warmup units
- Set Tenant::activate_now in `get_active_tenant_with_timeout` (the page
service's path for getting a reference to a tenant).
- Wait for tenant activation in HTTP handlers for timeline creation and
deletion: like page service requests, these require an active tenant and
should prioritize activation if called.
## Problem
Various places in remote storage were not subject to a timeout (thereby
stuck TCP connections could hold things up), and did not respect a
cancellation token (so things like timeline deletion or tenant detach
would have to wait arbitrarily long).
## Summary of changes
- Add download_cancellable and upload_cancellable helpers, and use them
in all the places we wait for remote storage operations (with the
exception of initdb downloads, where it would not have been safe).
- Add a cancellation token arg to `download_retry`.
- Use cancellation token args in various places that were missing one
per #5066Closes: #5066
Why is this only "basic" handling?
- Doesn't express difference between shutdown and errors in return
types, to avoid refactoring all the places that use an anyhow::Error
(these should all eventually return a more structured error type)
- Implements timeouts on top of remote storage, rather than within it:
this means that operations hitting their timeout will lose their
semaphore permit and thereby go to the back of the queue for their
retry.
- Doing a nicer job is tracked in
https://github.com/neondatabase/neon/issues/6096
Part of getpage@lsn benchmark epic:
https://github.com/neondatabase/neon/issues/5771
This PR moves the control plane's spread-all-over-the-place client for
the pageserver management API into a separate module within the
pageserver crate.
I need that client to be async in my benchmarking work, so, this PR
switches to the async version of `reqwest`.
That is also the right direction generally IMO.
The switch to async in turn mandated converting most of the
`control_plane/` code to async.
Note that some of the client methods should be taking `TenantShardId`
instead of `TenantId`, but, none of the callers seem to be
sharding-aware.
Leaving that for another time:
https://github.com/neondatabase/neon/issues/6154
* initdb uploads had no cancellation token, which means that when we
were stuck in upload retries, we wouldn't be able to delete the
timeline. in general, the combination of retrying forever and not having
cancellation tokens is quite dangerous.
* initdb uploads wouldn't rewind the file. this wasn't discovered in the
purposefully unreliable test-s3 in pytest because those fail on the
first byte always, not somewhere during the connection. we'd be getting
errors from the AWS sdk that the file was at an unexpected end.
slack thread: https://neondb.slack.com/archives/C033RQ5SPDH/p1702632247784079
Compaction was holding back timeline deletion because the compaction
lock had been acquired, but the semaphore was waited on. Timeline
deletion was waiting on the same lock for 1500s.
This replaces the
`pageserver::tenant::tasks::concurrent_background_tasks_rate_limit`
(which looks correct) with a simpler `..._permit` which is just an
infallible acquire, which is easier to spot "aah this needs to be raced
with cancellation tokens".
Ref: https://neondb.slack.com/archives/C03F5SM1N02/p1702496912904719
Ref: https://neondb.slack.com/archives/C03F5SM1N02/p1702578093497779
Before any json parsing from the http api only returned errors were per
field errors. Now they are done using `serde_path_to_error`, which at
least helped greatly with the `disk_usage_eviction_run` used for
testing. I don't think this can conflict with anything added in #5310.
## Problem
Historically, the pageserver used an "uninit mark" file on disk for two
purposes:
- Track which timeline dirs are incomplete for handling on restart
- Avoid trying to create the same timeline twice at the same time.
The original purpose of handling restarts is now defunct, as we use
remote storage as the source of truth and clean up any trash timeline
dirs on startup. Using the file to mutually exclude creation operations
is error prone compared with just doing it in memory, and the existing
checks happened some way into the creation operation, and could expose
errors as 500s (anyhow::Errors) rather than something clean.
## Summary of changes
- Creations are now mutually excluded in memory (using
`Tenant::timelines_creating`), rather than relying on a file on disk for
coordination.
- Acquiring unique access to the timeline ID now happens earlier in the
request.
- Creating the same timeline which already exists is now a 201: this
simplifies retry handling for clients.
- 409 is still returned if a timeline with the same ID is still being
created: if this happens it is probably because the client timed out an
earlier request and has retried.
- Colliding timeline creation requests should no longer return 500
errors
This paves the way to entirely removing uninit markers in a subsequent
change.
---------
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
## Problem
The websockets gauge for active db connections seems to be growing more
than the gauge for client connections over websockets, which does not
make sense.
## Summary of changes
refactor how our counter-pair gauges are represented. not sure if this
will improve the problem, but it should be harder to mess-up the
counters. The API is much nicer though now and doesn't require
scopeguard::defer hacks
Dependency (commits inline):
https://github.com/neondatabase/neon/pull/5842
## Problem
Secondary mode tenants need a manifest of what to download. Ultimately
this will be some kind of heat-scored set of layers, but as a robust
first step we will simply use the set of resident layers: secondary
tenant locations will aim to match the on-disk content of the attached
location.
## Summary of changes
- Add heatmap types representing the remote structure
- Add hooks to Tenant/Timeline for generating these heatmaps
- Create a new `HeatmapUploader` type that is external to `Tenant`, and
responsible for walking the list of attached tenants and scheduling
heatmap uploads.
Notes to reviewers:
- Putting the logic for uploads (and later, secondary mode downloads)
outside of `Tenant` is an opinionated choice, motivated by:
- Enable future smarter scheduling of operations, e.g. uploading the
stalest tenant first, rather than having all tenants compete for a fair
semaphore on a first-come-first-served basis. Similarly for downloads,
we may wish to schedule the tenants with the hottest un-downloaded
layers first.
- Enable accessing upload-related state without synchronization (it
belongs to HeatmapUploader, rather than being some Mutex<>'d part of
Tenant)
- Avoid further expanding the scope of Tenant/Timeline types, which are
already among the largest in the codebase
- You might reasonably wonder how much of the uploader code could be a
generic job manager thing. Probably some of it: but let's defer pulling
that out until we have at least two users (perhaps secondary downloads
will be the second one) to highlight which bits are really generic.
Compromises:
- Later, instead of using digests of heatmaps to decide whether anything
changed, I would prefer to avoid walking the layers in tenants that
don't have changes: tracking that will be a bit invasive, as it needs
input from both remote_timeline_client and Layer.
