# TLDR
All changes are no-op except
1. publishing additional metrics.
2. problem VI
## Problem I
It has come to my attention that the Neon Storage Controller doesn't
correctly update its "observed" state of tenants previously associated
with PSs that has come back up after a local data loss. It would still
think that the old tenants are still attached to page servers and won't
ask more questions. The pageserver has enough information from the
reattach request/response to tell that something is wrong, but it
doesn't do anything about it either. We need to detect this situation in
production while I work on a fix.
(I think there is just some misunderstanding about how Neon manages
their pageserver deployments which got me confused about all the
invariants.)
## Summary of changes I
Added a `pageserver_local_data_loss_suspected` gauge metric that will be
set to 1 if we detect a problematic situation from the reattch response.
The problematic situation is when the PS doesn't have any local tenants
but received a reattach response containing tenants.
We can set up an alert using this metric. The alert should be raised
whenever this metric reports non-zero number.
Also added a HTTP PUT
`http://pageserver/hadron-internal/reset_alert_gauges` API on the
pageserver that can be used to reset the gauge and the alert once we
manually rectify the situation (by restarting the HCC).
## Problem II
Azure upload is 3x slower than AWS. -> 3x slower ingestion.
The reason for the slower upload is that Azure upload in page server is
much slower => higher flush latency => higher disk consistent LSN =>
higher back pressure.
## Summary of changes II
Use Azure put_block API to uploads a 1 GB layer file in 8 blocks in
parallel.
I set the put_block block size to be 128 MB by default in azure config.
To minimize neon changes, upload function passes the layer file path to
the azure upload code through the storage metadata. This allows the
azure put block to use FileChunkStreamRead to stream read from one
partition in the file instead of loading all file data in memory and
split it into 8 128 MB chunks.
## How is this tested? II
1. rust test_real_azure tests the put_block change.
3. I deployed the change in azure dev and saw flush latency reduces from
~30 seconds to 10 seconds.
4. I also did a bunch of stress test using sqlsmith and 100 GB TPCDS
runs.
## Problem III
Currently Neon limits the compaction tasks as 3/4 * CPU cores. This
limits the overall compaction throughput and it can easily cause
head-of-the-line blocking problems when a few large tenants are
compacting.
## Summary of changes III
This PR increases the limit of compaction tasks as `BG_TASKS_PER_THREAD`
(default 4) * CPU cores. Note that `CONCURRENT_BACKGROUND_TASKS` also
limits some other tasks `logical_size_calculation` and `layer eviction`
. But compaction should be the most frequent and time-consuming task.
## Summary of changes IV
This PR adds the following PageServer metrics:
1. `pageserver_disk_usage_based_eviction_evicted_bytes_total`: captures
the total amount of bytes evicted. It's more straightforward to see the
bytes directly instead of layers.
2. `pageserver_active_storage_operations_count`: captures the active
storage operation, e.g., flush, L0 compaction, image creation etc. It's
useful to visualize these active operations to get a better idea of what
PageServers are spending cycles on in the background.
## Summary of changes V
When investigating data corruptions, it's useful to search the base
image and all WAL records of a page up to an LSN, i.e., a breakdown of
GetPage@LSN request. This PR implements this functionality with two
tools:
1. Extended `pagectl` with a new command to search the layer files for a
given key up to a given LSN from the `index_part.json` file. The output
can be used to download the files from S3 and then search the file
contents using the second tool.
Example usage:
```
cargo run --bin pagectl index-part search --tenant-id 09b99ea3239bbb3b2d883a59f087659d --timeline-id 7bedf4a6995baff7c0421ff9aebbcdab --path ~/Downloads/corruption/index_part.json-0000000c-formatted --key 000000067F000080140000802100000D61BD --lsn 70C/BF3D61D8
```
Example output:
```
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F0000801400008028000002FEFF__000007089F0B5381-0000070C7679EEB9-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000000000000000000000000000000000-000000067F0000801400008028000002F3F1__000006DD95B6F609-000006E2BA14C369-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F000080140000802100001B0973__000006D33429F539-000006DD95B6F609-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000164D81__000006C6343B2D31-000006D33429F539-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F0000801400008021000017687B__000006BA344FA7F1-000006C6343B2D31-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000165BAB__000006AD34613D19-000006BA344FA7F1-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000137A39__0000069F34773461-000006AD34613D19-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F000080140000802100000D4000-000000067F000080140000802100000F0000__0000069F34773460-0000000b
```
2. Added a unit test to search the layer file contents. It's not
implemented part of `pagectl` because it depends on some test harness
code, which can only be used by unit tests.
Example usage:
```
cargo test --package pageserver --lib -- tenant::debug::test_search_key --exact --nocapture -- --tenant-id 09b99ea3239bbb3b2d883a59f087659d --timeline-id 7bedf4a6995baff7c0421ff9aebbcdab --data-dir /Users/chen.luo/Downloads/corruption --key 000000067F000080140000802100000D61BD --lsn 70C/BF3D61D8
```
Example output:
```
# omitted image for brievity
delta: 69F/769D8180: will_init: false, "OgAAALGkuwXwYp12nwYAAECGAAASIqLHAAAAAH8GAAAUgAAAIYAAAL1hDQD/DLGkuwUDAAAAEAAWAA=="
delta: 69F/769CB6D8: will_init: false, "PQAAALGkuwXotZx2nwYAABAJAAAFk7tpACAGAH8GAAAUgAAAIYAAAL1hDQD/CQUAEAASALExuwUBAAAAAA=="
```
## Problem VI
Currently when page service resolves shards from page numbers, it
doesn't fully support the case that the shard could be split in the
middle. This will lead to query failures during the tenant split for
either commit or abort cases (it's mostly for abort).
## Summary of changes VI
This PR adds retry logic in `Cache::get()` to deal with shard resolution
errors more gracefully. Specifically, it'll clear the cache and retry,
instead of failing the query immediately. It also reduces the internal
timeout to make retries faster.
The PR also fixes a very obvious bug in
`TenantManager::resolve_attached_shard` where the code tries to cache
the computed the shard number, but forgot to recompute when the shard
count is different.
---------
Co-authored-by: William Huang <william.huang@databricks.com>
Co-authored-by: Haoyu Huang <haoyu.huang@databricks.com>
Co-authored-by: Chen Luo <chen.luo@databricks.com>
Co-authored-by: Vlad Lazar <vlad.lazar@databricks.com>
Co-authored-by: Vlad Lazar <vlad@neon.tech>
## Problem
https://github.com/neondatabase/neon/pull/11494 changes the batching
logic, but we don't have a way to evaluate it.
## Summary of changes
This PR introduces a global and per timeline metric which tracks the
reason for
which a batch was broken.
## Problem
We don't have metrics to exactly quantify the end user impact of
on-demand downloads.
Perf tracing is underway (#11140) to supply us with high-resolution
*samples*.
But it will also be useful to have some aggregate per-timeline and
per-instance metrics that definitively contain all observations.
## Summary of changes
This PR consists of independent commits that should be reviewed
independently.
However, for convenience, we're going to merge them together.
- refactor(metrics): measure_remote_op can use async traits
- impr(pageserver metrics): task_kind dimension for
remote_timeline_client latency histo
- implements https://github.com/neondatabase/cloud/issues/26800
- refs
https://github.com/neondatabase/cloud/issues/26193#issuecomment-2769705793
- use the opportunity to rename the metric and add a _global suffix;
checked grafana export, it's only used in two personal dashboards, one
of them mine, the other by Heikki
- log on-demand download latency for expensive-to-query but precise
ground truth
- metric for wall clock time spent waiting for on-demand downloads
## Refs
- refs https://github.com/neondatabase/cloud/issues/26800
- a bunch of minor investigations / incidents into latency outliers
## Problem
`TYPE_CHECKING` is used inconsistently across Python tests.
## Summary of changes
- Update `ruff`: 0.7.0 -> 0.11.2
- Enable TC (flake8-type-checking):
https://docs.astral.sh/ruff/rules/#flake8-type-checking-tc
- (auto)fix all new issues
## Problem
Previously, L0 flushes would wait for uploads, as a simple form of
backpressure. However, this prevented flush pipelining and upload
parallelism. It has since been disabled by default and replaced by L0
compaction backpressure.
Touches https://github.com/neondatabase/cloud/issues/24664.
## Summary of changes
This patch removes L0 flush upload waits, along with the
`l0_flush_wait_upload`. This can't be merged until the setting has been
removed across the fleet.
# Problem
We leave too few observability breadcrumbs in the case where wait_lsn is
exceptionally slow.
# Changes
- refactor: extract the monitoring logic out of `log_slow` into
`monitor_slow_future`
- add global + per-timeline counter for time spent waiting for wait_lsn
- It is updated while we're still waiting, similar to what we do for
page_service response flush.
- add per-timeline counterpair for started & finished wait_lsn count
- add slow-logging to leave breadcrumbs in logs, not just metrics
For the slow-logging, we need to consider not flooding the logs during a
broker or network outage/blip.
The solution is a "log-streak-level" concurrency limit per timeline.
At any given time, there is at most one slow wait_lsn that is logging
the "still running" and "completed" sequence of logs.
Other concurrent slow wait_lsn's don't log at all.
This leaves at least one breadcrumb in each timeline's logs if some
wait_lsn was exceptionally slow during a given period.
The full degree of slowness can then be determined by looking at the
per-timeline metric.
# Performance
Reran the `bench_log_slow` benchmark, no difference, so, existing call
sites are fine.
We do use a Semaphore, but only try_acquire it _after_ things have
already been determined to be slow. So, no baseline overhead
anticipated.
# Refs
-
https://github.com/neondatabase/cloud/issues/23486#issuecomment-2711587222
## Problem
We don't have per-timeline observability for read amplification.
Touches https://github.com/neondatabase/cloud/issues/23283.
## Summary of changes
Add a per-timeline `pageserver_layers_per_read` histogram.
NB: per-timeline histograms are expensive, but probably worth it in this
case.
This reverts commit 9e55d79803.
We'll still need this until we can tune L0 flush backpressure and
compaction. I'll add a setting to disable this separately.
This reapplies #10135. Just removing this flush backpressure without
further mitigations caused read amp increases during bulk ingestion
(predictably), so it was reverted. We will replace it by
compaction-based backpressure.
## Problem
In #8550, we made the flush loop wait for uploads after every layer.
This was to avoid unbounded buildup of uploads, and to reduce compaction
debt. However, the approach has several problems:
* It prevents upload parallelism.
* It prevents flush and upload pipelining.
* It slows down ingestion even when there is no need to backpressure.
* It does not directly backpressure based on compaction debt and read
amplification.
We will instead implement compaction-based backpressure in a PR
immediately following this removal (#5415).
Touches #5415.
Touches #10095.
## Summary of changes
Remove waiting on the upload queue in the flush loop.
## Refs
- Epic: https://github.com/neondatabase/neon/issues/9378
Co-authored-by: Vlad Lazar <vlad@neon.tech>
Co-authored-by: Christian Schwarz <christian@neon.tech>
## Problem
The read path does its IOs sequentially.
This means that if N values need to be read to reconstruct a page,
we will do N IOs and getpage latency is `O(N*IoLatency)`.
## Solution
With this PR we gain the ability to issue IO concurrently within one
layer visit **and** to move on to the next layer without waiting for IOs
from the previous visit to complete.
This is an evolved version of the work done at the Lisbon hackathon,
cf https://github.com/neondatabase/neon/pull/9002.
## Design
### `will_init` now sourced from disk btree index keys
On the algorithmic level, the only change is that the
`get_values_reconstruct_data`
now sources `will_init` from the disk btree index key (which is
PS-page_cache'd), instead
of from the `Value`, which is only available after the IO completes.
### Concurrent IOs, Submission & Completion
To separate IO submission from waiting for its completion, while
simultaneously
feature-gating the change, we introduce the notion of an `IoConcurrency`
struct
through which IO futures are "spawned".
An IO is an opaque future, and waiting for completions is handled
through
`tokio::sync::oneshot` channels.
The oneshot Receiver's take the place of the `img` and `records` fields
inside `VectoredValueReconstructState`.
When we're done visiting all the layers and submitting all the IOs along
the way
we concurrently `collect_pending_ios` for each value, which means
for each value there is a future that awaits all the oneshot receivers
and then calls into walredo to reconstruct the page image.
Walredo is now invoked concurrently for each value instead of
sequentially.
Walredo itself remains unchanged.
The spawned IO futures are driven to completion by a sidecar tokio task
that
is separate from the task that performs all the layer visiting and
spawning of IOs.
That tasks receives the IO futures via an unbounded mpsc channel and
drives them to completion inside a `FuturedUnordered`.
(The behavior from before this PR is available through
`IoConcurrency::Sequential`,
which awaits the IO futures in place, without "spawning" or "submitting"
them
anywhere.)
#### Alternatives Explored
A few words on the rationale behind having a sidecar *task* and what
alternatives were considered.
One option is to queue up all IO futures in a FuturesUnordered that is
polled
the first time when we `collect_pending_ios`.
Firstly, the IO futures are opaque, compiler-generated futures that need
to be polled at least once to submit their IO. "At least once" because
tokio-epoll-uring may not be able to submit the IO to the kernel on
first
poll right away.
Second, there are deadlocks if we don't drive the IO futures to
completion
independently of the spawning task.
The reason is that both the IO futures and the spawning task may hold
some
_and_ try to acquire _more_ shared limited resources.
For example, both spawning task and IO future may try to acquire
* a VirtualFile file descriptor cache slot async mutex (observed during
impl)
* a tokio-epoll-uring submission slot (observed during impl)
* a PageCache slot (currently this is not the case but we may move more
code into the IO futures in the future)
Another option is to spawn a short-lived `tokio::task` for each IO
future.
We implemented and benchmarked it during development, but found little
throughput improvement and moderate mean & tail latency degradation.
Concerns about pressure on the tokio scheduler made us discard this
variant.
The sidecar task could be obsoleted if the IOs were not arbitrary code
but a well-defined struct.
However,
1. the opaque futures approach taken in this PR allows leaving the
existing
code unchanged, which
2. allows us to implement the `IoConcurrency::Sequential` mode for
feature-gating
the change.
Once the new mode sidecar task implementation is rolled out everywhere,
and `::Sequential` removed, we can think about a descriptive submission
& completion interface.
The problems around deadlocks pointed out earlier will need to be solved
then.
For example, we could eliminate VirtualFile file descriptor cache and
tokio-epoll-uring slots.
The latter has been drafted in
https://github.com/neondatabase/tokio-epoll-uring/pull/63.
See the lengthy doc comment on `spawn_io()` for more details.
### Error handling
There are two error classes during reconstruct data retrieval:
* traversal errors: index lookup, move to next layer, and the like
* value read IO errors
A traversal error fails the entire get_vectored request, as before this
PR.
A value read error only fails that value.
In any case, we preserve the existing behavior that once
`get_vectored` returns, all IOs are done. Panics and failing
to poll `get_vectored` to completion will leave the IOs dangling,
which is safe but shouldn't happen, and so, a rate-limited
log statement will be emitted at warning level.
There is a doc comment on `collect_pending_ios` giving more code-level
details and rationale.
### Feature Gating
The new behavior is opt-in via pageserver config.
The `Sequential` mode is the default.
The only significant change in `Sequential` mode compared to before
this PR is the buffering of results in the `oneshot`s.
## Code-Level Changes
Prep work:
* Make `GateGuard` clonable.
Core Feature:
* Traversal code: track `will_init` in `BlobMeta` and source it from
the Delta/Image/InMemory layer index, instead of determining `will_init`
after we've read the value. This avoids having to read the value to
determine whether traversal can stop.
* Introduce `IoConcurrency` & its sidecar task.
* `IoConcurrency` is the clonable handle.
* It connects to the sidecar task via an `mpsc`.
* Plumb through `IoConcurrency` from high level code to the
individual layer implementations' `get_values_reconstruct_data`.
We piggy-back on the `ValuesReconstructState` for this.
* The sidecar task should be long-lived, so, `IoConcurrency` needs
to be rooted up "high" in the call stack.
* Roots as of this PR:
* `page_service`: outside of pagestream loop
* `create_image_layers`: when it is called
* `basebackup`(only auxfiles + replorigin + SLRU segments)
* Code with no roots that uses `IoConcurrency::sequential`
* any `Timeline::get` call
* `collect_keyspace` is a good example
* follow-up: https://github.com/neondatabase/neon/issues/10460
* `TimelineAdaptor` code used by the compaction simulator, unused in
practive
* `ingest_xlog_dbase_create`
* Transform Delta/Image/InMemoryLayer to
* do their values IO in a distinct `async {}` block
* extend the residence of the Delta/Image layer until the IO is done
* buffer their results in a `oneshot` channel instead of straight
in `ValuesReconstructState`
* the `oneshot` channel is wrapped in `OnDiskValueIo` /
`OnDiskValueIoWaiter`
types that aid in expressiveness and are used to keep track of
in-flight IOs so we can print warnings if we leave them dangling.
* Change `ValuesReconstructState` to hold the receiving end of the
`oneshot` channel aka `OnDiskValueIoWaiter`.
* Change `get_vectored_impl` to `collect_pending_ios` and issue walredo
concurrently, in a `FuturesUnordered`.
Testing / Benchmarking:
* Support queue-depth in pagebench for manual benchmarkinng.
* Add test suite support for setting concurrency mode ps config
field via a) an env var and b) via NeonEnvBuilder.
* Hacky helper to have sidecar-based IoConcurrency in tests.
This will be cleaned up later.
More benchmarking will happen post-merge in nightly benchmarks, plus in
staging/pre-prod.
Some intermediate helpers for manual benchmarking have been preserved in
https://github.com/neondatabase/neon/pull/10466 and will be landed in
later PRs.
(L0 layer stack generator!)
Drive-By:
* test suite actually didn't enable batching by default because
`config.compatibility_neon_binpath` is always Truthy in our CI
environment
=> https://neondb.slack.com/archives/C059ZC138NR/p1737490501941309
* initial logical size calculation wasn't always polled to completion,
which was
surfaced through the added WARN logs emitted when dropping a
`ValuesReconstructState` that still has inflight IOs.
* remove the timing histograms
`pageserver_getpage_get_reconstruct_data_seconds`
and `pageserver_getpage_reconstruct_seconds` because with planning,
value read
IO, and walredo happening concurrently, one can no longer attribute
latency
to any one of them; we'll revisit this when Vlad's work on
tracing/sampling
through RequestContext lands.
* remove code related to `get_cached_lsn()`.
The logic around this has been dead at runtime for a long time,
ever since the removal of the materialized page cache in #8105.
## Testing
Unit tests use the sidecar task by default and run both modes in CI.
Python regression tests and benchmarks also use the sidecar task by
default.
We'll test more in staging and possibly preprod.
# Future Work
Please refer to the parent epic for the full plan.
The next step will be to fold the plumbing of IoConcurrency
into RequestContext so that the function signatures get cleaned up.
Once `Sequential` isn't used anymore, we can take the next
big leap which is replacing the opaque IOs with structs
that have well-defined semantics.
---------
Co-authored-by: Christian Schwarz <christian@neon.tech>
## Problem
In #8550, we made the flush loop wait for uploads after every layer.
This was to avoid unbounded buildup of uploads, and to reduce compaction
debt. However, the approach has several problems:
* It prevents upload parallelism.
* It prevents flush and upload pipelining.
* It slows down ingestion even when there is no need to backpressure.
* It does not directly backpressure WAL ingestion (only via
`disk_consistent_lsn`), and will build up in-memory layers.
* It does not directly backpressure based on compaction debt and read
amplification.
An alternative solution to these problems is proposed in #8390.
In the meanwhile, we revert the change to reduce the impact on ingest
throughput. This does reintroduce some risk of unbounded
upload/compaction buildup. Until
https://github.com/neondatabase/neon/issues/8390, this can be addressed
in other ways:
* Use `max_replication_apply_lag` (aka `remote_consistent_lsn`), which
will more directly limit upload debt.
* Shard the tenant, which will spread the flush/upload work across more
Pageservers and move the bottleneck to Safekeeper.
Touches #10095.
## Summary of changes
Remove waiting on the upload queue in the flush loop.
## Problem
With pipelining enabled, the time a request spends in the batcher stage
counts towards the smgr op latency.
If pipelining is disabled, that time is not accounted for.
In practice, this results in a jump in smgr getpage latencies in various
dashboards and degrades the internal SLO.
## Solution
In a similar vein to #10042 and with a similar rationale, this PR stops
counting the time spent in batcher stage towards smgr op latency.
The smgr op latency metric is reduced to the actual execution time.
Time spent in batcher stage is tracked in a separate histogram.
I expect to remove that histogram after batching rollout is complete,
but it will be helpful in the meantime to reason about the rollout.
## Problem
In #9962 I changed the smgr metrics to include time spent on flush.
It isn't under our (=storage team's) control how long that flush takes
because the client can stop reading requests.
## Summary of changes
Stop the timer as soon as we've buffered up the response in the
`pgb_writer`.
Track flush time in a separate metric.
---------
Co-authored-by: Yuchen Liang <70461588+yliang412@users.noreply.github.com>
## Problem
With the current metrics we can't identify which shards are ingesting
data at any given time.
## Summary of changes
Add a metric for the number of wal records received for processing by
each shard. This is per (tenant, timeline, shard).
## Problem
There's no metrics for disk consistent LSN and remote LSN. This stuff is
useful when looking at ingest performance.
## Summary of changes
Two per timeline metrics are added: `pageserver_disk_consistent_lsn` and
`pageserver_projected_remote_consistent_lsn`. I went for the projected
remote lsn instead of the visible one
because that more closely matches remote storage write tput. Ideally we
would have both, but these metrics are expensive.
This PR
- fixes smgr metrics https://github.com/neondatabase/neon/issues/9925
- adds an additional startup log line logging the current batching
config
- adds a histogram of batch sizes global and per-tenant
- adds a metric exposing the current batching config
The issue described #9925 is that before this PR, request latency was
only observed *after* batching.
This means that smgr latency metrics (most importantly getpage latency)
don't account for
- `wait_lsn` time
- time spent waiting for batch to fill up / the executor stage to pick
up the batch.
The fix is to use a per-request batching timer, like we did before the
initial batching PR.
We funnel those timers through the entire request lifecycle.
I noticed that even before the initial batching changes, we weren't
accounting for the time spent writing & flushing the response to the
wire.
This PR drive-by fixes that deficiency by dropping the timers at the
very end of processing the batch, i.e., after the `pgb.flush()` call.
I was **unable to maintain the behavior that we deduct
time-spent-in-throttle from various latency metrics.
The reason is that we're using a *single* counter in `RequestContext` to
track micros spent in throttle.
But there are *N* metrics timers in the batch, one per request.
As a consequence, the practice of consuming the counter in the drop
handler of each timer no longer works because all but the first timer
will encounter error `close() called on closed state`.
A failed attempt to maintain the current behavior can be found in
https://github.com/neondatabase/neon/pull/9951.
So, this PR remvoes the deduction behavior from all metrics.
I started a discussion on Slack about it the implications this has for
our internal SLO calculation:
https://neondb.slack.com/archives/C033RQ5SPDH/p1732910861704029
# Refs
- fixes https://github.com/neondatabase/neon/issues/9925
- sub-issue https://github.com/neondatabase/neon/issues/9377
- epic: https://github.com/neondatabase/neon/issues/9376
## Problem
On Debian 12 (Bookworm), Python 3.11 is the latest available version.
## Summary of changes
- Update Python to 3.11 in build-tools
- Fix ruff check / format
- Fix mypy
- Use `StrEnum` instead of pair `str`, `Enum`
- Update docs
In complement to
https://github.com/neondatabase/tokio-epoll-uring/pull/56.
## Problem
We want to make tokio-epoll-uring slots waiters queue depth observable
via Prometheus.
## Summary of changes
- Add `pageserver_tokio_epoll_uring_slots_submission_queue_depth`
metrics as a `Histogram`.
- Each thread-local tokio-epoll-uring system is given a `LocalHistogram`
to observe the metrics.
- Keep a list of `Arc<ThreadLocalMetrics>` used on-demand to flush data
to the shared histogram.
- Extend `Collector::collect` to report
`pageserver_tokio_epoll_uring_slots_submission_queue_depth`.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
Co-authored-by: Christian Schwarz <christian@neon.tech>
## Problem
We recently added a "visibility" state to layers, but nothing
initializes it.
Part of:
- #8398
## Summary of changes
- Add a dependency on `range-set-blaze`, which is used as a fast
incrementally updated alternative to KeySpace. We could also use this to
replace the internals of KeySpaceRandomAccum if we wanted to. Writing a
type that does this kind of "BtreeMap & merge overlapping entries" thing
isn't super complicated, but no reason to write this ourselves when
there's a third party impl available.
- Add a function to layermap to calculate visibilities for each layer
- Add a function to Timeline to call into layermap and then apply these
visibilities to the Layer objects.
- Invoke the calculation during startup, after image layer creations,
and when removing branches. Branch removal and image layer creation are
the two ways that a layer can go from Visible to Covered.
- Add unit test & benchmark for the visibility calculation
- Expose `pageserver_visible_physical_size` metric, which should always
be <= `pageserver_remote_physical_size`.
- This metric will feed into the /v1/utilization endpoint later: the
visible size indicates how much space we would like to use on this
pageserver for this tenant.
- When `pageserver_visible_physical_size` is greater than
`pageserver_resident_physical_size`, this is a sign that the tenant has
long-idle branches, which result in layers that are visible in
principle, but not used in practice.
This does not keep visibility hints up to date in all cases:
particularly, when creating a child timeline, any previously covered
layers will not get marked Visible until they are accessed.
Updates after image layer creation could be implemented as more of a
special case, but this would require more new code: the existing depth
calculation code doesn't maintain+yield the list of deltas that would be
covered by an image layer.
## Performance
This operation is done rarely (at startup and at timeline deletion), so
needs to be efficient but not ultra-fast.
There is a new `visibility` bench that measures runtime for a synthetic
100k layers case (`sequential`) and a real layer map (`real_map`) with
~26k layers.
The benchmark shows runtimes of single digit milliseconds (on a ryzen
7950). This confirms that the runtime shouldn't be a problem at startup
(as we already incur S3-level latencies there), but that it's slow
enough that we definitely shouldn't call it more often than necessary,
and it may be worthwhile to optimize further later (things like: when
removing a branch, only bother scanning layers below the branchpoint)
```
visibility/sequential time: [4.5087 ms 4.5894 ms 4.6775 ms]
change: [+2.0826% +3.9097% +5.8995%] (p = 0.00 < 0.05)
Performance has regressed.
Found 24 outliers among 100 measurements (24.00%)
2 (2.00%) high mild
22 (22.00%) high severe
min: 0/1696070, max: 93/1C0887F0
visibility/real_map time: [7.0796 ms 7.0832 ms 7.0871 ms]
change: [+0.3900% +0.4505% +0.5164%] (p = 0.00 < 0.05)
Change within noise threshold.
Found 4 outliers among 100 measurements (4.00%)
3 (3.00%) high mild
1 (1.00%) high severe
min: 0/1696070, max: 93/1C0887F0
visibility/real_map_many_branches
time: [4.5285 ms 4.5355 ms 4.5434 ms]
change: [-1.0012% -0.8004% -0.5969%] (p = 0.00 < 0.05)
Change within noise threshold.
```
## Problem
We lack insight into:
- How much of a tenant's physical size is image vs. delta layers
- Average sizes of image vs. delta layers
- Total layer counts per timeline, indicating size of index_part object
As well as general observability love, this is motivated by
https://github.com/neondatabase/neon/issues/6738, where we need to
define some sensible thresholds for storage amplification, and using
total physical size may not work well (if someone does a lot of DROPs
then it's legitimate for the physical-synthetic ratio to be huge), but
the ratio between image layer size and delta layer size may be a better
indicator of whether we're generating unreasonable quantities of image
layers.
## Summary of changes
- Add pageserver_layer_bytes and pageserver_layer_count metrics,
labelled by timeline and `kind` (delta or image)
- Add & subtract these with LayerInner's lifetime.
I'm intentionally avoiding using a generic metric RAII guard object, to
avoid bloating LayerInner: it already has all the information it needs
to update metric on new+drop.
## Problem
The metrics we have today aren't convenient for planning around the
impact of timeline archival on costs.
Closes: https://github.com/neondatabase/neon/issues/8108
## Summary of changes
- Add metric `pageserver_archive_size`, which indicates the logical
bytes of data which we would expect to write into an archived branch.
- Add metric `pageserver_pitr_history_size`, which indicates the
distance between last_record_lsn and the PITR cutoff.
These metrics are somewhat temporary: when we implement #8088 and
associated consumption metric changes, these will reach a final form.
For now, an "archived" branch is just any branch outside of its parent's
PITR window: later, archival will become an explicit state (which will
_usually_ correspond to falling outside the parent's PITR window).
The overall volume of timeline metrics is something to watch, but we are
removing many more in https://github.com/neondatabase/neon/pull/8245
than this PR is adding.
Part of #7497, closes#8120.
## Summary of changes
This PR adds a metric to track the number of valid leases after `GCInfo`
gets refreshed each time.
Besides this metric, we should also track disk space and synthetic size
(after #8071 is closed) to make sure leases are used properly.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
part of Epic https://github.com/neondatabase/neon/issues/7386
# Motivation
The materialized page cache adds complexity to the code base, which
increases the maintenance burden and risk for subtle and hard to
reproduce bugs such as #8050.
Further, the best hit rate that we currently achieve in production is ca
1% of materialized page cache lookups for
`task_kind=PageRequestHandler`. Other task kinds have hit rates <0.2%.
Last, caching page images in Pageserver rewards under-sized caches in
Computes because reading from Pageserver's materialized page cache over
the network is often sufficiently fast (low hundreds of microseconds).
Such Computes should upscale their local caches to fit their working
set, rather than repeatedly requesting the same page from Pageserver.
Some more discussion and context in internal thread
https://neondb.slack.com/archives/C033RQ5SPDH/p1718714037708459
# Changes
This PR removes the materialized page cache code & metrics.
The infrastructure for different key kinds in `PageCache` is left in
place, even though the "Immutable" key kind is the only remaining one.
This can be further simplified in a future commit.
Some tests started failing because their total runtime was dependent on
high materialized page cache hit rates. This test makes them
fixed-runtime or raises pytest timeouts:
* test_local_file_cache_unlink
* test_physical_replication
* test_pg_regress
# Performance
I focussed on ensuring that this PR will not result in a performance
regression in prod.
* **getpage** requests: our production metrics have shown the
materialized page cache to be irrelevant (low hit rate). Also,
Pageserver is the wrong place to cache page images, it should happen in
compute.
* **ingest** (`task_kind=WalReceiverConnectionHandler`): prod metrics
show 0 percent hit rate, so, removing will not be a regression.
* **get_lsn_by_timestamp**: important API for branch creation, used by
control pane. The clog pages that this code uses are not
materialize-page-cached because they're not 8k. No risk of introducing a
regression here.
We will watch the various nightly benchmarks closely for more results
before shipping to prod.
ref https://github.com/neondatabase/neon/issues/7443
## Summary of changes
This pull request adds a size estimator for aux files. Each timeline
stores a cached `isize` for the estimated total size of aux files. It
gets reset on basebackup, and gets updated for each aux file
modification. TODO: print a warning when it exceeds the size.
The size metrics is not accurate. Race between `on_basebackup` and other
functions could create a negative basebackup size, but the chance is
rare. Anyways, this does not impose any extra I/Os to the storage as
everything is computed in-memory.
The aux files are only stored on shard 0. As basebackups are only
generated on shard 0, only shard 0 will report this metrics.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
We recently went through an incident where compaction was inhibited by a
bug. We didn't observe this until quite late because we did not have alerting
on deep reads.
## Summary of changes
+ Tweak an existing metric that tracks the depth of a read on the
non-vectored read path:
* Give it a better name
* Track all layers
* Larger buckets
+ Add a similar metric for the vectored read path
+ Add a compaction smoke test which uses these metrics. This test would
have caught
the compaction issue mentioned earlier.
Related https://github.com/neondatabase/neon/issues/7428
refs #6737
# Problem
Before this PR, on-demand downloads weren't measured per tenant_id.
This makes root-cause analysis of latency spikes harder, requiring us to
resort to log scraping for
```
{neon_service="pageserver"} |= `downloading on-demand` |= `$tenant_id`
```
which can be expensive when zooming out in Grafana.
Context: https://neondb.slack.com/archives/C033RQ5SPDH/p1707809037868189
# Solution / Changes
- Remove the calls_started histogram
- I did the dilegence, there are only 2 dashboards using this histogram,
and in fact only one uses it as a histogram, the other just as a
a counter.
- [Link
1](8115b54d9f/neonprod/dashboards/hkXNF7oVz/dashboard-Z31XmM24k.yaml (L1454)):
`Pageserver Thrashing` dashboard, linked from playbook, will fix.
- [Link
2](8115b54d9f/neonprod/dashboards/CEllzAO4z/dashboard-sJqfNFL4k.yaml (L599)):
one of my personal dashboards, unused for a long time, already broken in
other ways, no need to fix.
- replace `pageserver_remote_timeline_client_calls_unfinished` gauge
with a counter pair
- Required `Clone`-able `IntCounterPair`, made the necessary changes in
the `libs/metrics` crate
- fix tests to deal with the fallout
A subsequent PR will remove a timeline-scoped metric to compensate.
Note that we don't need additional global counters for the per-timeline
counters affected by this PR; we can use the `remote_storage` histogram
for those, which, conveniently, also include the secondary-mode
downloads, which aren't covered by the remote timeline client metrics
(should they?).
## Problem
When investigating test failures
(https://github.com/neondatabase/neon/issues/6813) I noticed we were
doing a bunch of Reconciler runs right after splitting a tenant.
It's because the splitting test does a pageserver restart, and there was
a bug in /re-attach handling, where we would update the generation
correctly in the database and intent state, but not observed state,
thereby triggering a reconciliation on the next call to maybe_reconcile.
This didn't break anything profound (underlying rules about generations
were respected), but caused the storage controller to do an un-needed
extra round of bumping the generation and reconciling.
## Summary of changes
- Start adding metrics to the storage controller
- Assert on the number of reconciles done in test_sharding_split_smoke
- Fix /re-attach to update `observed` such that we don't spuriously
re-reconcile tenants.
There is O(n^2) issues due to how we store these directories (#6626), so
it's good to keep an eye on them and ensure the numbers stay low.
The new per-timeline metric `pageserver_directory_entries_count`
isn't perfect, namely we don't calculate it every time we attach
the timeline, but only if there is an actual change.
Also, it is a collective metric over multiple scalars. Lastly,
we only emit the metric if it is above a certain threshold.
However, the metric still give a feel for the general size of the timeline.
We care less for small values as the metric is mainly there to
detect and track tenants with large directory counts.
We also expose the directory counts in `TimelineInfo` so that one can
get the detailed size distribution directly via the pageserver's API.
Related: #6642 , https://github.com/neondatabase/cloud/issues/10273
Before tenant migration it made sense to leak broken tenants in the
metrics until restart. Nowdays it makes less sense because on
cancellations we set the tenant broken. The set metric still allows
filterable alerting.
Fixes: #6507
## Problem
To test sharding, we need something to control it. We could write python
code for doing this from the test runner, but this wouldn't be usable
with neon_local run directly, and when we want to write tests with large
number of shards/tenants, Rust is a better fit efficiently handling all
the required state.
This service enables automated tests to easily get a system with
sharding/HA without the test itself having to set this all up by hand:
existing tests can be run against sharded tenants just by setting a
shard count when creating the tenant.
## Summary of changes
Attachment service was previously a map of TenantId->TenantState, where
the principal state stored for each tenant was the generation and the
last attached pageserver. This enabled it to serve the re-attach and
validate requests that the pageserver requires.
In this PR, the scope of the service is extended substantially to do
overall management of tenants in the pageserver, including
tenant/timeline creation, live migration, evacuation of offline
pageservers etc. This is done using synchronous code to make declarative
changes to the tenant's intended state (`TenantState.policy` and
`TenantState.intent`), which are then translated into calls into the
pageserver by the `Reconciler`.
Top level summary of modules within
`control_plane/attachment_service/src`:
- `tenant_state`: structure that represents one tenant shard.
- `service`: implements the main high level such as tenant/timeline
creation, marking a node offline, etc.
- `scheduler`: for operations that need to pick a pageserver for a
tenant, construct a scheduler and call into it.
- `compute_hook`: receive notifications when a tenant shard is attached
somewhere new. Once we have locations for all the shards in a tenant,
emit an update to postgres configuration via the neon_local `LocalEnv`.
- `http`: HTTP stubs. These mostly map to methods on `Service`, but are
separated for readability and so that it'll be easier to adapt if/when
we switch to another RPC layer.
- `node`: structure that describes a pageserver node. The most important
attribute of a node is its availability: marking a node offline causes
tenant shards to reschedule away from it.
This PR is a precursor to implementing the full sharding service for
prod (#6342). What's the difference between this and a production-ready
controller for pageservers?
- JSON file persistence to be replaced with a database
- Limited observability.
- No concurrency limits. Marking a pageserver offline will try and
migrate every tenant to a new pageserver concurrently, even if there are
thousands.
- Very simple scheduler that only knows to pick the pageserver with
fewest tenants, and place secondary locations on a different pageserver
than attached locations: it does not try to place shards for the same
tenant on different pageservers. This matters little in tests, because
picking the least-used pageserver usually results in round-robin
placement.
- Scheduler state is rebuilt exhaustively for each operation that
requires a scheduler.
- Relies on neon_local mechanisms for updating postgres: in production
this would be something that flows through the real control plane.
---------
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
When doing global queries in VictoriaMetrics, the per-timeline
histograms make us run into cardinality limits.
We don't want to give them up just yet because we don't
have an alternative for drilling down on timeline-specific
performance issues.
So, add a pre-aggregated histogram and add observations to it
whenever we add observations to the per-timeline histogram.
While we're at it, switch to using a strummed enum for the operation
type names.
We currently have a timeseries for each of the tenants in different
states. We only want this for Broken. Other states could be counters.
Fix this by making the `pageserver_tenant_states_count` a counter
without a `tenant_id` and
add a `pageserver_broken_tenants_count` which has a `tenant_id` label,
each broken tenant being 1.
The histogram distinguishes by ok/err.
I took the liberty to create a small abstraction for such use cases.
It helps keep the label values inside `metrics.rs`, right next
to the place where the metric and its labels are declared.
## Problem
Some metrics are better to be observed at page-server level. Otherwise,
as we have a lot of tenants in production, we cannot do a sum b/c
Prometheus has limit on how many time series we can aggregate. This also
helps reduce metrics scraping size.
## Summary of changes
Some integration tests are likely not to pass as it will check the
existence of some metrics. Waiting for CI complete and fix them.
Metrics downgraded: page cache hit (where we are likely to have a
page-server level page cache in the future instead of per-tenant), and
reconstruct time (this would better be tenant-level, as we have one pg
replayer for each tenant, but now we make it page-server level as we do
not need that fine-grained data).
---------
Signed-off-by: Alex Chi <iskyzh@gmail.com>
## Problem
This PR includes doc changes to the current metrics as well as adding
new metrics. With the new set of metrics, we can quantitatively analyze
the read amp., write amp. and space amp. in the system, when used
together with https://github.com/neondatabase/neonbench
close https://github.com/neondatabase/neon/issues/4312
ref https://github.com/neondatabase/neon/issues/4347
compaction metrics TBD, a novel idea is to print L0 file number and
number of layers in the system, and we can do this in the future when we
start working on compaction.
## Summary of changes
* Add `READ_NUM_FS_LAYERS` for computing read amp.
* Add `MATERIALIZED_PAGE_CACHE_HIT_UPON_REQUEST`.
* Add `GET_RECONSTRUCT_DATA_TIME`. GET_RECONSTRUCT_DATA_TIME +
RECONSTRUCT_TIME + WAIT_LSN_TIME should be approximately total time of
reads.
* Add `5.0` and `10.0` to `STORAGE_IO_TIME_BUCKETS` given some fsync
runs slow (i.e., > 1s) in some cases.
* Some `WAL_REDO` metrics are only used when Postgres is involved in the
redo process.
---------
Signed-off-by: Alex Chi <iskyzh@gmail.com>
