This makes it possible for the compiler to validate that a match block
matched all PostgreSQL versions we support.
## Problem
We did not have a complete picture about which places we had to test
against PG versions, and what format these versions were: The full PG
version ID format (Major/minor/bugfix `MMmmbb`) as transfered in
protocol messages, or only the Major release version (`MM`). This meant
type confusion was rampant.
With this change, it becomes easier to develop new version-dependent
features, by making type and niche confusion impossible.
## Summary of changes
Every use of `pg_version` is now typed as either `PgVersionId` (u32,
valued in decimal `MMmmbb`) or PgMajorVersion (an enum, with a value for
every major version we support, serialized and stored like a u32 with
the value of that major version)
---------
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
Previously, we couldn't read from an in-memory layer while a batch was
being written to it. Vice-versa, we couldn't write to it while there
was an on-going read.
## Summary of Changes
The goal of this change is to improve concurrency. Writes happened
through a &mut self method so the enforcement was at the type system
level.
We attempt to improve by:
1. Adding interior mutability to EphemeralLayer. This involves wrapping
the buffered writer in a read-write lock.
2. Minimise the time that the read lock is held for. Only hold the read
lock while reading from the buffers (recently flushed or pending
flush). If we need to read from the file, drop the lock and allow IO
to be concurrent.
The new benchmark variants with concurrent reads improve between 70 to
200 percent (against main).
Benchmark results are in this
[commit](891f094ce6).
## Future Changes
We can push the interior mutability into the buffered writer. The
mutable tail goes under a read lock, the flushed part goes into an
ArcSwap and then we can read from anything that is flushed _without_ any
locking.
Introduce a separate `postgres_ffi_types` crate which contains a few
types and functions that were used in the API. `postgres_ffi_types` is a
much small crate than `postgres_ffi`, and it doesn't depend on bindgen
or the Postgres C headers.
Move NeonWalRecord and Value types to wal_decoder crate. They are only
used in the pageserver-safekeeper "ingest" API. The rest of the ingest
API types are defined in wal_decoder, so move these there as well.
The expected operating range for the production NVMe drives is
in the range of 50 to 250us.
The bucket boundaries before this PR were not well suited
to reason about the utilization / queuing / latency variability
of those devices.
# Performance
There was some concern about perf impact of having so many buckets,
considering the impl does a linear search on each observe().
I added a benchmark and measured on relevant machines.
In any way, the PR is 40 buckets, so, won't make a meaningful
difference on production machines (im4gn.2xlarge),
going from 30ns -> 35ns.
This PR adds a runtime validation mode to check adherence to alignment
and size-multiple requirements at the VirtualFile level.
This can help prevent alignment bugs from slipping into production
because test systems may have more lax requirements than production.
(This is not the case today, but it could change in the future).
It also allows catching O_DIRECT bugs on systems that don't have
O_DIRECT (macOS).
Consequently, we can now accept
`virtual_file_io_mode={direct,direct-rw}` on macOS now.
This has the side benefit of removing some annoying conditional
compilation around `IoMode`.
A third benefit is that it helped weed out size-multiple requirement
violation bugs in how the VirtualFile unit tests exercise read and write
APIs.
I seized the opportunity to trim these tests down to what actually
matters, i.e., exercising of the `OpenFiles` file descriptor cache.
Lastly, this PR flips the binary-built-in default to `DirectRw` so that
when running Python regress tests and benchmarks without specifying
`PAGESERVER_VIRTUAL_FILE_IO_MODE`, one gets the production behavior.
Refs
- fixes https://github.com/neondatabase/neon/issues/11676
# Problem
The Pageserver read path exclusively uses direct IO if
`virtual_file_io_mode=direct`.
The write path is half-finished. Here is what the various writing
components use:
|what|buffering|flags on <br/>`v_f_io_mode`<br/>=`buffered`|flags on
<br/>`virtual_file_io_mode`<br/>=`direct`|
|-|-|-|-|
|`DeltaLayerWriter`| BlobWriter<BUFFERED=true> | () | () |
|`ImageLayerWriter`| BlobWriter<BUFFERED=false> | () | () |
|`download_layer_file`|BufferedWriter|()|()|
|`InMemoryLayer`|BufferedWriter|()|O_DIRECT|
The vehicle towards direct IO support is `BufferedWriter` which
- largely takes care of O_DIRECT alignment & size-multiple requirements
- double-buffering to mask latency
`DeltaLayerWriter`, `ImageLayerWriter` use `blob_io::BlobWriter` , which
has neither of these.
# Changes
## High-Level
At a high-level this PR makes the following primary changes:
- switch the two layer writer types to use `BufferedWriter` & make
sensitive to `virtual_file_io_mode` (via open_with_options_**v2**)
- make `download_layer_file` sensitive to `virtual_file_io_mode` (also
via open_with_options_**v2**)
- add `virtual_file_io_mode=direct-rw` as a feature gate
- we're hackish-ly piggybacking on OpenOptions's ask for write access
here
- this means with just `=direct` InMemoryLayer reads and writes no
longer uses O_DIRECT
- this is transitory and we'll remove the `direct-rw` variant once the
rollout is complete
(The `_v2` APIs for opening / creating VirtualFile are those that are
sensitive to `virtual_file_io_mode`)
The result is:
|what|uses <br/>`BufferedWriter`|flags on
<br/>`v_f_io_mode`<br/>=`buffered`|flags on
<br/>`v_f_io_mode`<br/>=`direct`|flags on
<br/>`v_f_io_mode`<br/>=`direct-rw`|
|-|-|-|-|-|
|`DeltaLayerWriter`| ~~Blob~~BufferedWriter | () | () | O_DIRECT |
|`ImageLayerWriter`| ~~Blob~~BufferedWriter | () | () | O_DIRECT |
|`download_layer_file`|BufferedWriter|()|()|O_DIRECT|
|`InMemoryLayer`|BufferedWriter|()|~~O_DIRECT~~()|O_DIRECT|
## Code-Level
The main change is:
- Switch `blob_io::BlobWriter` away from its own buffering method to use
`BufferedWriter`.
Additional prep for upholding `O_DIRECT` requirements:
- Layer writer `finish()` methods switched to use IoBufferMut for
guaranteed buffer address alignment. The size of the buffers is PAGE_SZ
and thereby implicitly assumed to fulfill O_DIRECT requirements.
For the hacky feature-gating via `=direct-rw`:
- Track `OpenOptions::write(true|false)` in a field; bunch of mechanical
churn.
- Consolidate the APIs in which we "open" or "create" VirtualFile for
better overview over which parts of the code use the `_v2` APIs.
Necessary refactorings & infra work:
- Add doc comments explaining how BufferedWriter ensures that writes are
compliant with O_DIRECT alignment & size constraints. This isn't new,
but should be spelled out.
- Add the concept of shutdown modes to `BufferedWriter::shutdown` to
make writer shutdown adhere to these constraints.
- The `PadThenTruncate` mode might not be necessary in practice because
I believe all layer files ever written are sized in multiples `PAGE_SZ`
and since `PAGE_SZ` is larger than the current alignment requirements
(512/4k depending on platform), it won't be necesary to pad.
- Some test (I believe `round_trip_test_compressed`?) required it though
- [ ] TODO: decide if we want to accept that complexity; if we do then
address TODO in the code to separate alignment requirement from buffer
capacity
- Add `set_len` (=`ftruncate`) VirtualFile operation to support the
above.
- Allow `BufferedWriter` to start at a non-zero offset (to make room for
the summary block).
Cleanups unlocked by this change:
- Remove non-positional APIs from VirtualFile (e.g. seek, write_full,
read_full)
Drive-by fixes:
- PR https://github.com/neondatabase/neon/pull/11585 aimed to run unit
tests for all `virtual_file_io_mode` combinations but didn't because of
a missing `_` in the env var.
# Performance
This section assesses this PR's impact on deployments with current
production setting (`=direct`) and anticipated impact of switching to
(`=direct-rw`).
For `DeltaLayerWriter`, `=direct` should remain unchanged to slightly
improved on throughput because the `BlobWriter`'s buffer had the same
size as the `BufferedWriter`'s buffer, but it didn't have the
double-buffering that `BufferedWriter` has.
The `=direct-rw` enables direct IO; throughput should not be suffering
because of double-buffering; benchmarks will show if this is true.
The `ImageLayerWriter` was previously not doing any buffering
(`BUFFERED=false`).
It went straight to issuing the IO operation to the underlying
VirtualFile and the buffering was done by the kernel.
The switch to `BufferedWriter` under `=direct` adds an additional memcpy
into the BufferedWriter's buffer.
We will win back that memcpy when enabling direct IO via `=direct-rw`.
A nice win from the switch to `BufferedWriter` is that ImageLayerWriter
performs >=16x fewer write operations to VirtualFile (the BlobWriter
performs one write per len field and one write per image value).
This should save low tens of microseconds of CPU overhead from doing all
these syscalls/io_uring operations, regardless of `=direct` or
`=direct-rw`.
Aside from problems with alignment, this write frequency without
double-buffering is prohibitive if we actually have to wait for the
disk, which is what will happen when we enable direct IO via
(`=direct-rw`).
Throughput should not be suffering because of BufferedWrite's
double-buffering; benchmarks will show if this is true.
`InMemoryLayer` at `=direct` will flip back to using buffered IO but
remain on BufferedWriter.
The buffered IO adds back one memcpy of CPU overhead.
Throughput should not suffer and will might improve on
not-memory-pressured Pageservers but let's remember that we're doing the
whole direct IO thing to eliminate global memory pressure as a source of
perf variability.
## bench_ingest
I reran `bench_ingest` on `im4gn.2xlarge` and `Hetzner AX102`.
Use `git diff` with `--word-diff` or similar to see the change.
General guidance on interpretation:
- immediate production impact of this PR without production config
change can be gauged by comparing the same `io_mode=Direct`
- end state of production switched over to `io_mode=DirectRw` can be
gauged by comparing old results' `io_mode=Direct` to new results'
`io_mode=DirectRw`
Given above guidance, on `im4gn.2xlarge`
- immediate impact is a significant improvement in all cases
- end state after switching has same significant improvements in all
cases
- ... except `ingest/io_mode=DirectRw volume_mib=128 key_size_bytes=8192
key_layout=Sequential write_delta=Yes` which only achieves `238 MiB/s`
instead of `253.43 MiB/s`
- this is a 6% degradation
- this workload is typical for image layer creation
# Refs
- epic https://github.com/neondatabase/neon/issues/9868
- stacked atop
- preliminary refactor https://github.com/neondatabase/neon/pull/11549
- bench_ingest overhaul https://github.com/neondatabase/neon/pull/11667
- derived from https://github.com/neondatabase/neon/pull/10063
Co-authored-by: Yuchen Liang <yuchen@neon.tech>
In #10063 we will switch BlobWriter to use the owned buffers IO buffered
writer, which implements double-buffering by virtue of a background task
that performs the flushing.
That task's lifecylce must be contained within the Timeline lifecycle,
so, it must hold the timeline gate open and respect Timeline::cancel.
This PR does the noisy plumbing to reduce the #10063 diff.
Refs
- extracted from https://github.com/neondatabase/neon/pull/10063
- epic https://github.com/neondatabase/neon/issues/9868
## Problem
Not a complete fix for https://github.com/neondatabase/neon/issues/11492
but should work for a short term.
Our current retry strategy for walredo is to retry every request exactly
once. This retry doesn't make sense because it retries all requests
exactly once and each error is expected to cause process restart and
cause future requests to fail. I'll explain it with a scenario of two
threads requesting redos: one with an invalid history (that will cause
walredo to panic) and another that has a correct redo sequence.
First let's look at how we handle retries right now in
do_with_walredo_process. At the beginning of the function it will spawn
a new process if there's no existing one. Then it will continue to redo.
If the process fails, the first process that encounters the error will
remove the walredo process object from the OnceCell, so that the next
time it gets accessed, a new process will be spawned; if it is the last
one that uses the old walredo process, it will kill and wait the process
in `drop(proc)`. I'm skeptical whether this works under races but I
think this is not the root cause of the problem. In this retry handler,
if there are N requests attached to a walredo process and the i-th
request fails (panics the walredo), all other N-i requests will fail and
they need to retry so that they can access a new walredo process.
```
time ---->
proc A None B
request 1 ^-----------------^ fail
uses A for redo replace with None
request 2 ^-------------------- fail
uses A for redo
request 3 ^----------------^ fail
uses A for redo last ref, wait for A to be killed
request 4 ^---------------
None, spawn new process B
```
The problem is with our retry strategy. Normally, for a system that we
want to retry on, the probability of errors for each of the requests are
uncorrelated. However, in walredo, a prior request that panics the
walredo process will cause all future walredo on that process to fail
(that's correlated).
So, back to the situation where we have 2 requests where one will
definitely fail and the other will succeed and we get the following
sequence, where retry attempts = 1,
* new walredo process A starts.
* request 1 (invalid) being processed on A and panics A, waiting for
retry, remove process A from the process object.
* request 2 (valid) being processed on A and receives pipe broken /
poisoned process error, waiting for retry, wait for A to be killed --
this very likely takes a while and cannot finish before request 1 gets
processed again
* new walredo process B starts.
* request 1 (invalid) being processed again on B and panics B, the whole
request fail.
* request 2 (valid) being processed again on B, and get a poisoned error
again.
```
time ---->
proc A None B None
request 1 ^-----------------^--------------^--------------------^
spawn A for redo fail spawn B for redo fail
request 2 ^--------------------^-------------------------^------------^
use A for redo fail, wait to kill A B for redo fail again
```
In such cases, no matter how we set n_attempts, as long as the retry
count applies to all requests, this sequence is bound to fail both
requests because of how they get sequenced; while we could potentially
make request 2 successful.
There are many solutions to this -- like having a separate walredo
manager for compactions, or define which errors are retryable (i.e.,
broken pipe can be retried, while real walredo error won't be retried),
or having a exclusive big lock over the whole redo process (the current
one is very fine-grained). In this patch, we go with a simple approach:
use different retry attempts for different types of requests.
For gc-compaction, the attempt count is set to 0, so that it never
retries and consequently stops the compaction process -- no more redo
will be issued from gc-compaction. Once the walredo process gets
restarted, the normal read requests will proceed normally.
## Summary of changes
Add redo_attempt for each reconstruct value request to set different
retry policies.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Erik Grinaker <erik@neon.tech>
In
-
https://github.com/neondatabase/neon/pull/10993#issuecomment-2690428336
I added infinite retries for buffered writer flush IOs, primarily to
gracefully handle ENOSPC but more generally so that the buffered writer
is not left in a state where reads from the surrounding InMemoryLayer
cause panics.
However, I didn't add cancellation sensitivity, which is concerning
because then there is no way to detach a timeline/tenant that is
encountering the write IO errors.
That’s a legitimate scenario in the case of some edge case bug.
See the #10993 description for details.
This PR
- first makes flush loop infallible, enabled by infinite retries
- then adds sensitivity to `Timeline::cancel` to the flush loop, thereby
making it fallible in one specific way again
- finally fixes the InMemoryLayer/EphemeralFile/BufferedWriter
amalgamate to remain read-available after flush loop is cancelled.
The support for read-availability after cancellation is necessary so
that reads from the InMemoryLayer that are already queued up behind the
RwLock that wraps the BufferedWriter won't panic because of the
`mutable=None` that we leave behind in case the flush loop gets
cancelled.
# Alternatives
One might think that we can only ship the change for read-availability
if flush encounters an error, without the infinite retrying and/or
cancellation sensitivity complexity.
The problem with that is that read-availability sounds good but is
really quite useless, because we cannot ingest new WAL without a
writable InMemoryLayer. Thus, very soon after we transition to read-only
mode, reads from compute are going to wait anyway, but on `wait_lsn`
instead of the RwLock, because ingest isn't progressing.
Thus, having the infinite flush retries still makes more sense because
they're just "slowness" to the user, whereas wait_lsn is hard errors.
We use the `metrics` / `prometheus` crate in the Pageserver code base.
This PR demonstrates
- typical performance pitfalls with that crate
- our current set of techniques to avoid most of these pitfalls.
refs
- https://github.com/neondatabase/neon/issues/10948
- https://github.com/neondatabase/neon/pull/7202
- I applied the `label_values__cache_label_values_lookup` technique
there.
- It didn't yield measurable results in high-level benchmarks though.
## Problem
Image layers may be nested inside in-memory layers as diagnosed
[here](https://github.com/neondatabase/neon/issues/10720#issuecomment-2649419252).
The read path doesn't support this and may skip over the image layer,
resulting in a failure to reconstruct the page.
## Summary of changes
We already support nesting of image layers inside delta layers. The
logic lives in `LayerMap::select_layer`.
The main goal of this PR is to propagate the candidate in-memory layer
down to that point and update
the selection logic.
Important changes are:
1. Support partial reads for the in-memory layer. Previously, we could
only specify the start LSN of the read.
We need to control the end LSN too.
2. `LayerMap::ranged_search` considers in-memory layers too. Previously,
the search for in-memory layers
was done explicitly in `Timeline::get_reconstruct_data_timeline`. Note
that `LayerMap::ranged_search` now returns
a weak readable layer which the `LayerManager` can upgrade. This dance
is such that we can unit test the layer selection logic.
3. Update `LayerMap::select_layer` to consider the candidate in-memory
layer too
Loosely related drive bys:
1. Remove the "keys not found" tracking in the ranged search. This
wasn't used anywhere and it just complicates things.
2. Remove the difficulty map stuff from the layer map. Again, not used
anywhere.
Closes https://github.com/neondatabase/neon/issues/9185
Closes https://github.com/neondatabase/neon/issues/10720
Updates storage components to edition 2024. We like to stay on the
latest edition if possible. There is no functional changes, however some
code changes had to be done to accommodate the edition's breaking
changes.
The PR has two commits:
* the first commit updates storage crates to edition 2024 and appeases
`cargo clippy` by changing code. i have accidentially ran the formatter
on some files that had other edits.
* the second commit performs a `cargo fmt`
I would recommend a closer review of the first commit and a less close
review of the second one (as it just runs `cargo fmt`).
part of https://github.com/neondatabase/neon/issues/10918
## Problem
With upload queue reordering in #10218, we can easily get into a
situation where multiple index uploads are queued back to back, which
can't be parallelized. This will happen e.g. when multiple layer flushes
enqueue layer/index/layer/index/... and the layers skip the queue and
are uploaded in parallel.
These index uploads will incur serial S3 roundtrip latencies, and may
block later operations.
Touches #10096.
## Summary of changes
When multiple back-to-back index uploads are ready to upload, only
upload the most recent index and drop the rest.
## Problem
The upload queue can currently schedule an arbitrary number of tasks.
This can both spawn an unbounded number of Tokio tasks, and also
significantly slow down upload queue scheduling as it's quadratic in
number of operations.
Touches #10096.
## Summary of changes
Limit the number of inprogress tasks to the remote storage upload
concurrency. While this concurrency limit is shared across all tenants,
there's certainly no point in scheduling more than this -- we could even
consider setting the limit lower, but don't for now to avoid
artificially constraining tenants.
## Problem
The upload queue currently sees significant head-of-line blocking. For
example, index uploads act as upload barriers, and for every layer flush
we schedule a layer and index upload, which effectively serializes layer
uploads.
Resolves#10096.
## Summary of changes
Allow upload queue operations to bypass the queue if they don't conflict
with preceding operations, increasing parallelism.
NB: the upload queue currently schedules an explicit barrier after every
layer flush as well (see #8550). This must be removed to enable
parallelism. This will require a better mechanism for compaction
backpressure, see e.g. #8390 or #5415.
Closes#9387.
## Problem
`BufferedWriter` cannot proceed while the owned buffer is flushing to
disk. We want to implement double buffering so that the flush can happen
in the background. See #9387.
## Summary of changes
- Maintain two owned buffers in `BufferedWriter`.
- The writer is in charge of copying the data into owned, aligned
buffer, once full, submit it to the flush task.
- The flush background task is in charge of flushing the owned buffer to
disk, and returned the buffer to the writer for reuse.
- The writer and the flush background task communicate through a
bi-directional channel.
For in-memory layer, we also need to be able to read from the buffered
writer in `get_values_reconstruct_data`. To handle this case, we did the
following
- Use replace `VirtualFile::write_all` with `VirtualFile::write_all_at`,
and use `Arc` to share it between writer and background task.
- leverage `IoBufferMut::freeze` to get a cheaply clonable `IoBuffer`,
one clone will be submitted to the channel, the other clone will be
saved within the writer to serve reads. When we want to reuse the
buffer, we can invoke `IoBuffer::into_mut`, which gives us back the
mutable aligned buffer.
- InMemoryLayer reads is now aware of the maybe_flushed part of the
buffer.
**Caveat**
- We removed the owned version of write, because this interface does not
work well with buffer alignment. The result is that without direct IO
enabled,
[`download_object`](a439d57050/pageserver/src/tenant/remote_timeline_client/download.rs (L243))
does one more memcpy than before this PR due to the switch to use
`_borrowed` version of the write.
- "Bypass aligned part of write" could be implemented later to avoid
large amount of memcpy.
**Testing**
- use an oneshot channel based control mechanism to make flush behavior
deterministic in test.
- test reading from `EphemeralFile` when the last submitted buffer is
not flushed, in-progress, and done flushing to disk.
## Performance
We see performance improvement for small values, and regression on big
values, likely due to being CPU bound + disk write latency.
[Results](https://www.notion.so/neondatabase/Benchmarking-New-BufferedWriter-11-20-2024-143f189e0047805ba99acda89f984d51?pvs=4)
## Checklist before requesting a review
- [ ] I have performed a self-review of my code.
- [ ] If it is a core feature, I have added thorough tests.
- [ ] Do we need to implement analytics? if so did you add the relevant
metrics to the dashboard?
- [ ] If this PR requires public announcement, mark it with
/release-notes label and add several sentences in this section.
## Checklist before merging
- [ ] Do not forget to reformat commit message to not include the above
checklist
---------
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
Co-authored-by: Christian Schwarz <christian@neon.tech>
## Problem
`no_sync` initially just skipped syncfs on startup (#9677). I'm also
interested in flaky tests that time out during pageserver shutdown while
flushing l0s, so to eliminate disk throughput as a source of issues
there,
## Summary of changes
- Drive-by change for test timeouts: add a couple more ::info logs
during pageserver startup so it's obvious which part got stuck.
- Add a SyncMode enum to configure VirtualFile and respect it in
sync_all and sync_data functions
- During pageserver startup, set SyncMode according to `no_sync`
## Problem
https://github.com/neondatabase/neon/pull/9524 split the decoding and
interpretation step from ingestion.
The output of the first phase is a `wal_decoder::models::InterpretedWalRecord`.
Before this patch set that struct contained a list of `Value` instances.
We wish to lift the decoding and interpretation step to the safekeeper,
but it would be nice if the safekeeper gave us a batch containing the raw data instead of actual values.
## Summary of changes
Main goal here is to make `InterpretedWalRecord` hold a raw buffer which
contains pre-serialized Values.
For this we do:
1. Add a `SerializedValueBatch` type. This is `inmemory_layer::SerializedBatch` with some
extra functionality for extension, observing values for shard 0 and tests.
2. Replace `inmemory_layer::SerializedBatch` with `SerializedValueBatch`
3. Make `DatadirModification` maintain a `SerializedValueBatch`.
### `DatadirModification` changes
`DatadirModification` now maintains a `SerializedValueBatch` and extends
it as new WAL records come in (to avoid flushing to disk on every
record).
In turn, this cascaded into a number of modifications to
`DatadirModification`:
1. Replace `pending_data_pages` and `pending_zero_data_pages` with `pending_data_batch`.
2. Removal of `pending_zero_data_pages` and its cousin `on_wal_record_end`
3. Rename `pending_bytes` to `pending_metadata_bytes` since this is what it tracks now.
4. Adapting of various utility methods like `len`, `approx_pending_bytes` and `has_dirty_data_pages`.
Removal of `pending_zero_data_pages` and the optimisation associated
with it ((1) and (2)) deserves more detail.
Previously all zero data pages went through `pending_zero_data_pages`.
We wrote zero data pages when filling gaps caused by relation extension
(case A) and when handling special wal records (case B). If it happened
that the same WAL record contained a non zero write for an entry in
`pending_zero_data_pages` we skipped the zero write.
Case A: We handle this differently now. When ingesting the
`SerialiezdValueBatch` associated with one PG WAL record, we identify the gaps and fill the
them in one go. Essentially, we move from a per key process (gaps were filled after each
new key), and replace it with a per record process. Hence, the optimisation is not
required anymore.
Case B: When the handling of a special record needs to zero out a key,
it just adds that to the current batch. I inspected the code, and I
don't think the optimisation kicked in here.
## Problem
We wish to have high level WAL decoding logic in `wal_decoder::decoder`
module.
## Summary of Changes
For this we need the `Value` and `NeonWalRecord` types accessible there, so:
1. Move `Value` and `NeonWalRecord` to `pageserver::value` and
`pageserver::record` respectively.
2. Get rid of `pageserver::repository` (follow up from (1))
3. Move PG specific WAL record types to `postgres_ffi::walrecord`. In
theory they could live in `wal_decoder`, but it would create a circular
dependency between `wal_decoder` and `postgres_ffi`. Long term it makes
sense for those types to be PG version specific, so that will work out nicely.
4. Move higher level WAL record types (to be ingested by pageserver)
into `wal_decoder::models`
Related: https://github.com/neondatabase/neon/issues/9335
Epic: https://github.com/neondatabase/neon/issues/9329
Part of #8130
## Problem
Pageserver previously goes through the kernel page cache for all the
IOs. The kernel page cache makes light-loaded pageserver have deceptive
fast performance. Using direct IO would offer predictable latencies of
our virtual file IO operations.
In particular for reads, the data pages also have an extremely low
temporal locality because the most frequently accessed pages are cached
on the compute side.
## Summary of changes
This PR enables pageserver to use direct IO for delta layer and image
layer reads. We can ship them separately because these layers are
write-once, read-many, so we will not be mixing buffered IO with direct
IO.
- implement `IoBufferMut`, an buffer type with aligned allocation
(currently set to 512).
- use `IoBufferMut` at all places we are doing reads on image + delta
layers.
- leverage Rust type system and use `IoBufAlignedMut` marker trait to
guarantee that the input buffers for the IO operations are aligned.
- page cache allocation is also made aligned.
_* in-memory layer reads and the write path will be shipped separately._
## Testing
Integration test suite run with O_DIRECT enabled:
https://github.com/neondatabase/neon/pull/9350
## Performance
We evaluated performance based on the `get-page-at-latest-lsn`
benchmark. The results demonstrate a decrease in the number of IOps, no
sigificant change in the latency mean, and an slight improvement on the
p99.9 and p99.99 latencies.
[Benchmark](https://www.notion.so/neondatabase/Benchmark-O_DIRECT-for-image-and-delta-layers-2024-10-01-112f189e00478092a195ea5a0137e706?pvs=4)
## Rollout
We will add `virtual_file_io_mode=direct` region by region to enable
direct IO on image + delta layers.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
## Problem
We need a way to incrementally switch to direct IO. During the rollout
we might want to switch to O_DIRECT on image and delta layer read path
first before others.
## Summary of changes
- Revisited and simplified direct io config in `PageserverConf`.
- We could add a fallback mode for open, but for read there isn't a
reasonable alternative (without creating another buffered virtual file).
- Added a wrapper around `VirtualFile`, current implementation become
`VirtualFileInner`
- Use `open_v2`, `create_v2`, `open_with_options_v2` when we want to use
the IO mode specified in PS config.
- Once we onboard all IO through VirtualFile using this new API, we will
delete the old code path.
- Make io mode live configurable for benchmarking.
- Only guaranteed for files opened after the config change, so do it
before the experiment.
As an example, we are using `open_v2` with
`virtual_file::IoMode::Direct` in
https://github.com/neondatabase/neon/pull/9169
We also remove `io_buffer_alignment` config in
a04cfd754b and use it as a compile time
constant. This way we don't have to carry the alignment around or make
frequent call to retrieve this information from the static variable.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
Not used in production, but in benchmarks, to demonstrate minimal RTT.
(It would be nice to not have to copy the 8KiB of zeroes, but, that
would require larger protocol changes).
Found this useful in investigation
https://github.com/neondatabase/neon/pull/8952.
This PR simplifies the pageserver configuration parsing as follows:
* introduce the `pageserver_api::config::ConfigToml` type
* implement `Default` for `ConfigToml`
* use serde derive to do the brain-dead leg-work of processing the toml
document
* use `serde(default)` to fill in default values
* in `pageserver` crate:
* use `toml_edit` to deserialize the pageserver.toml string into a
`ConfigToml`
* `PageServerConfig::parse_and_validate` then
* consumes the `ConfigToml`
* destructures it exhaustively into its constituent fields
* constructs the `PageServerConfig`
The rules are:
* in `ConfigToml`, use `deny_unknown_fields` everywhere
* static default values go in `pageserver_api`
* if there cannot be a static default value (e.g. which default IO
engine to use, because it depends on the runtime), make the field in
`ConfigToml` an `Option`
* if runtime-augmentation of a value is needed, do that in
`parse_and_validate`
* a good example is `virtual_file_io_engine` or `l0_flush`, both of
which need to execute code to determine the effective value in
`PageServerConf`
The benefits:
* massive amount of brain-dead repetitive code can be deleted
* "unused variable" compile-time errors when removing a config value,
due to the exhaustive destructuring in `parse_and_validate`
* compile-time errors guide you when adding a new config field
Drawbacks:
* serde derive is sometimes a bit too magical
* `deny_unknown_fields` is easy to miss
Future Work / Benefits:
* make `neon_local` use `pageserver_api` to construct `ConfigToml` and
write it to `pageserver.toml`
* This provides more type safety / coompile-time errors than the current
approach.
### Refs
Fixes#3682
### Future Work
* `remote_storage` deser doesn't reject unknown fields
https://github.com/neondatabase/neon/issues/8915
* clean up `libs/pageserver_api/src/config.rs` further
* break up into multiple files, at least for tenant config
* move `models` as appropriate / refine distinction between config and
API models / be explicit about when it's the same
* use `pub(crate)` visibility on `mod defaults` to detect stale values
Part of [Epic: Bypass PageCache for user data
blocks](https://github.com/neondatabase/neon/issues/7386).
# Problem
`InMemoryLayer` still uses the `PageCache` for all data stored in the
`VirtualFile` that underlies the `EphemeralFile`.
# Background
Before this PR, `EphemeralFile` is a fancy and (code-bloated) buffered
writer around a `VirtualFile` that supports `blob_io`.
The `InMemoryLayerInner::index` stores offsets into the `EphemeralFile`.
At those offset, we find a varint length followed by the serialized
`Value`.
Vectored reads (`get_values_reconstruct_data`) are not in fact vectored
- each `Value` that needs to be read is read sequentially.
The `will_init` bit of information which we use to early-exit the
`get_values_reconstruct_data` for a given key is stored in the
serialized `Value`, meaning we have to read & deserialize the `Value`
from the `EphemeralFile`.
The L0 flushing **also** needs to re-determine the `will_init` bit of
information, by deserializing each value during L0 flush.
# Changes
1. Store the value length and `will_init` information in the
`InMemoryLayer::index`. The `EphemeralFile` thus only needs to store the
values.
2. For `get_values_reconstruct_data`:
- Use the in-memory `index` figures out which values need to be read.
Having the `will_init` stored in the index enables us to do that.
- View the EphemeralFile as a byte array of "DIO chunks", each 512 bytes
in size (adjustable constant). A "DIO chunk" is the minimal unit that we
can read under direct IO.
- Figure out which chunks need to be read to retrieve the serialized
bytes for thes values we need to read.
- Coalesce chunk reads such that each DIO chunk is only read once to
serve all value reads that need data from that chunk.
- Merge adjacent chunk reads into larger
`EphemeralFile::read_exact_at_eof_ok` of up to 128k (adjustable
constant).
3. The new `EphemeralFile::read_exact_at_eof_ok` fills the IO buffer
from the underlying VirtualFile and/or its in-memory buffer.
4. The L0 flush code is changed to use the `index` directly, `blob_io`
5. We can remove the `ephemeral_file::page_caching` construct now.
The `get_values_reconstruct_data` changes seem like a bit overkill but
they are necessary so we issue the equivalent amount of read system
calls compared to before this PR where it was highly likely that even if
the first PageCache access was a miss, remaining reads within the same
`get_values_reconstruct_data` call from the same `EphemeralFile` page
were a hit.
The "DIO chunk" stuff is truly unnecessary for page cache bypass, but,
since we're working on [direct
IO](https://github.com/neondatabase/neon/issues/8130) and
https://github.com/neondatabase/neon/issues/8719 specifically, we need
to do _something_ like this anyways in the near future.
# Alternative Design
The original plan was to use the `vectored_blob_io` code it relies on
the invariant of Delta&Image layers that `index order == values order`.
Further, `vectored_blob_io` code's strategy for merging IOs is limited
to adjacent reads. However, with direct IO, there is another level of
merging that should be done, specifically, if multiple reads map to the
same "DIO chunk" (=alignment-requirement-sized and -aligned region of
the file), then it's "free" to read the chunk into an IO buffer and
serve the two reads from that buffer.
=> https://github.com/neondatabase/neon/issues/8719
# Testing / Performance
Correctness of the IO merging code is ensured by unit tests.
Additionally, minimal tests are added for the `EphemeralFile`
implementation and the bit-packed `InMemoryLayerIndexValue`.
Performance testing results are presented below.
All pref testing done on my M2 MacBook Pro, running a Linux VM.
It's a release build without `--features testing`.
We see definitive improvement in ingest performance microbenchmark and
an ad-hoc microbenchmark for getpage against InMemoryLayer.
```
baseline: commit 7c74112b2a origin/main
HEAD: ef1c55c52e
```
<details>
```
cargo bench --bench bench_ingest -- 'ingest 128MB/100b seq, no delta'
baseline
ingest-small-values/ingest 128MB/100b seq, no delta
time: [483.50 ms 498.73 ms 522.53 ms]
thrpt: [244.96 MiB/s 256.65 MiB/s 264.73 MiB/s]
HEAD
ingest-small-values/ingest 128MB/100b seq, no delta
time: [479.22 ms 482.92 ms 487.35 ms]
thrpt: [262.64 MiB/s 265.06 MiB/s 267.10 MiB/s]
```
</details>
We don't have a micro-benchmark for InMemoryLayer and it's quite
cumbersome to add one. So, I did manual testing in `neon_local`.
<details>
```
./target/release/neon_local stop
rm -rf .neon
./target/release/neon_local init
./target/release/neon_local start
./target/release/neon_local tenant create --set-default
./target/release/neon_local endpoint create foo
./target/release/neon_local endpoint start foo
psql 'postgresql://cloud_admin@127.0.0.1:55432/postgres'
psql (13.16 (Debian 13.16-0+deb11u1), server 15.7)
CREATE TABLE wal_test (
id SERIAL PRIMARY KEY,
data TEXT
);
DO $$
DECLARE
i INTEGER := 1;
BEGIN
WHILE i <= 500000 LOOP
INSERT INTO wal_test (data) VALUES ('data');
i := i + 1;
END LOOP;
END $$;
-- => result is one L0 from initdb and one 137M-sized ephemeral-2
DO $$
DECLARE
i INTEGER := 1;
random_id INTEGER;
random_record wal_test%ROWTYPE;
start_time TIMESTAMP := clock_timestamp();
selects_completed INTEGER := 0;
min_id INTEGER := 1; -- Minimum ID value
max_id INTEGER := 100000; -- Maximum ID value, based on your insert range
iters INTEGER := 100000000; -- Number of iterations to run
BEGIN
WHILE i <= iters LOOP
-- Generate a random ID within the known range
random_id := min_id + floor(random() * (max_id - min_id + 1))::int;
-- Select the row with the generated random ID
SELECT * INTO random_record
FROM wal_test
WHERE id = random_id;
-- Increment the select counter
selects_completed := selects_completed + 1;
-- Check if a second has passed
IF EXTRACT(EPOCH FROM clock_timestamp() - start_time) >= 1 THEN
-- Print the number of selects completed in the last second
RAISE NOTICE 'Selects completed in last second: %', selects_completed;
-- Reset counters for the next second
selects_completed := 0;
start_time := clock_timestamp();
END IF;
-- Increment the loop counter
i := i + 1;
END LOOP;
END $$;
./target/release/neon_local stop
baseline: commit 7c74112b2a origin/main
NOTICE: Selects completed in last second: 1864
NOTICE: Selects completed in last second: 1850
NOTICE: Selects completed in last second: 1851
NOTICE: Selects completed in last second: 1918
NOTICE: Selects completed in last second: 1911
NOTICE: Selects completed in last second: 1879
NOTICE: Selects completed in last second: 1858
NOTICE: Selects completed in last second: 1827
NOTICE: Selects completed in last second: 1933
ours
NOTICE: Selects completed in last second: 1915
NOTICE: Selects completed in last second: 1928
NOTICE: Selects completed in last second: 1913
NOTICE: Selects completed in last second: 1932
NOTICE: Selects completed in last second: 1846
NOTICE: Selects completed in last second: 1955
NOTICE: Selects completed in last second: 1991
NOTICE: Selects completed in last second: 1973
```
NB: the ephemeral file sizes differ by ca 1MiB, ours being 1MiB smaller.
</details>
# Rollout
This PR changes the code in-place and is not gated by a feature flag.
Part of #8130, closes#8719.
## Problem
Currently, vectored blob io only coalesce blocks if they are immediately
adjacent to each other. When we switch to Direct IO, we need a way to
coalesce blobs that are within the dio-aligned boundary but has gap
between them.
## Summary of changes
- Introduces a `VectoredReadCoalesceMode` for `VectoredReadPlanner` and
`StreamingVectoredReadPlanner` which has two modes:
- `AdjacentOnly` (current implementation)
- `Chunked(<alignment requirement>)`
- New `ChunkedVectorBuilder` that considers batching `dio-align`-sized
read, the start and end of the vectored read will respect
`stx_dio_offset_align` / `stx_dio_mem_align` (`vectored_read.start` and
`vectored_read.blobs_at.first().start_offset` will be two different
value).
- Since we break the assumption that blobs within single `VectoredRead`
are next to each other (implicit end offset), we start to store blob end
offsets in the `VectoredRead`.
- Adapted existing tests to run in both `VectoredReadCoalesceMode`.
- The io alignment can also be live configured at runtime.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
## Problem/Solution
TimelineWriter::put_batch is simply a loop over individual puts. Each
put acquires and releases locks, and checks for potentially starting a
new layer. Batching these is more efficient, but more importantly
unlocks future changes where we can pre-build serialized buffers much
earlier in the ingest process, potentially even on the safekeeper
(imagine a future model where some variant of DatadirModification lives
on the safekeeper).
Ensuring that the values in put_batch are written to one layer also
enables a simplification upstream, where we no longer need to write
values in LSN-order. This saves us a sort, but also simplifies follow-on
refactors to DatadirModification: we can store metadata keys and data
keys separately at that level without needing to zip them together in
LSN order later.
## Why?
In this PR, these changes are simplify optimizations, but they are
motivated by evolving the ingest path in the direction of disentangling
extracting DatadirModification from Timeline. It may not obvious how
right now, but the general idea is that we'll end up with three phases
of ingest:
- A) Decode walrecords and build a datadirmodification with all the
simple data contents already in a big serialized buffer ready to write
to an ephemeral layer **<-- this part can be pipelined and parallelized,
and done on a safekeeper!**
- B) Let that datadirmodification see a Timeline, so that it can also
generate all the metadata updates that require a read-modify-write of
existing pages
- C) Dump the results of B into an ephemeral layer.
Related: https://github.com/neondatabase/neon/issues/8452
## Caveats
Doing a big monolithic buffer of values to write to disk is ordinarily
an anti-pattern: we prefer nice streaming I/O. However:
- In future, when we do this first decode stage on the safekeeper, it
would be inefficient to serialize a Vec of Value, and then later
deserialize it just to add blob size headers while writing into the
ephemeral layer format. The idea is that for bulk write data, we will
serialize exactly once.
- The monolithic buffer is a stepping stone to pipelining more of this:
by seriailizing earlier (rather than at the final put_value), we will be
able to parallelize the wal decoding and bulk serialization of data page
writes.
- The ephemeral layer's buffered writer already stalls writes while it
waits to flush: so while yes we'll stall for a couple milliseconds to
write a couple megabytes, we already have stalls like this, just
distributed across smaller writes.
## Benchmarks
This PR is primarily a stepping stone to safekeeper ingest filtering,
but also provides a modest efficiency improvement to the `wal_recovery`
part of `test_bulk_ingest`.
test_bulk_ingest:
```
test_bulk_insert[neon-release-pg16].insert: 23.659 s
test_bulk_insert[neon-release-pg16].pageserver_writes: 5,428 MB
test_bulk_insert[neon-release-pg16].peak_mem: 626 MB
test_bulk_insert[neon-release-pg16].size: 0 MB
test_bulk_insert[neon-release-pg16].data_uploaded: 1,922 MB
test_bulk_insert[neon-release-pg16].num_files_uploaded: 8
test_bulk_insert[neon-release-pg16].wal_written: 1,382 MB
test_bulk_insert[neon-release-pg16].wal_recovery: 18.981 s
test_bulk_insert[neon-release-pg16].compaction: 0.055 s
vs. tip of main:
test_bulk_insert[neon-release-pg16].insert: 24.001 s
test_bulk_insert[neon-release-pg16].pageserver_writes: 5,428 MB
test_bulk_insert[neon-release-pg16].peak_mem: 604 MB
test_bulk_insert[neon-release-pg16].size: 0 MB
test_bulk_insert[neon-release-pg16].data_uploaded: 1,922 MB
test_bulk_insert[neon-release-pg16].num_files_uploaded: 8
test_bulk_insert[neon-release-pg16].wal_written: 1,382 MB
test_bulk_insert[neon-release-pg16].wal_recovery: 23.586 s
test_bulk_insert[neon-release-pg16].compaction: 0.054 s
```
## Problem
This follows a PR that insists all input keys are representable in 16
bytes:
- https://github.com/neondatabase/neon/pull/8648
& a PR that prevents postgres from sending us keys that use the high
bits of field2:
- https://github.com/neondatabase/neon/pull/8657
Motivation for this change:
1. Ingest is bottlenecked on CPU
2. InMemoryLayer can create huge (~1M value) BTreeMap<Key,_> for its
index.
3. Maps over i128 are much faster than maps over an arbitrary 18 byte
struct.
It may still be worthwhile to make the index two-tier to optimize for
the case where only the last 4 bytes (blkno) of the key vary frequently,
but simply using the i128 representation of keys has a big impact for
very little effort.
Related: #8452
## Summary of changes
- Introduce `CompactKey` type which contains an i128
- Use this instead of Key in InMemoryLayer's index, converting back and
forth as needed.
## Performance
All the small-value `bench_ingest` cases show improved throughput.
The one that exercises this index most directly shows a 35% throughput
increase:
```
ingest-small-values/ingest 128MB/100b seq, no delta
time: [374.29 ms 378.56 ms 383.38 ms]
thrpt: [333.88 MiB/s 338.13 MiB/s 341.98 MiB/s]
change:
time: [-26.993% -26.117% -25.111%] (p = 0.00 < 0.05)
thrpt: [+33.531% +35.349% +36.974%]
Performance has improved.
```
## Problem
Some developers build on MacOS, which doesn't have io_uring.
## Summary of changes
- Add `io_engine_for_bench`, which on linux will give io_uring or panic
if it's unavailable, and on MacOS will always panic.
We do not want to run such benchmarks with StdFs: the results aren't
interesting, and will actively waste the time of any developers who
start investigating performance before they realize they're using a
known-slow I/O backend.
Why not just conditionally compile this benchmark on linux only? Because
even on linux, I still want it to refuse to run if it can't get
io_uring.
Ephemeral files cleanup on drop but did not delay shutdown, leading to
problems with restarting the tenant. The solution is as proposed:
- make ephemeral files carry the gate guard to delay `Timeline::gate`
closing
- flush in-memory layers and strong references to those on
`Timeline::shutdown`
The above are realized by making LayerManager an `enum` with `Open` and
`Closed` variants, and fail requests to modify `LayerMap`.
Additionally:
- fix too eager anyhow conversions in compaction
- unify how we freeze layers and handle errors
- optimize likely_resident_layers to read LayerFileManager hashmap
values instead of bouncing through LayerMap
Fixes: #7830
## Problem
We lack a rust bench for the inmemory layer and delta layer write paths:
it is useful to benchmark these components independent of postgres & WAL
decoding.
Related: https://github.com/neondatabase/neon/issues/8452
## Summary of changes
- Refactor DeltaLayerWriter to avoid carrying a Timeline, so that it can
be cleanly tested + benched without a Tenant/Timeline test harness. It
only needed the Timeline for building `Layer`, so this can be done in a
separate step.
- Add `bench_ingest`, which exercises a variety of workload "shapes"
(big values, small values, sequential keys, random keys)
- Include a small uncontroversial optimization: in `freeze`, only
exhaustively walk values to assert ordering relative to end_lsn in debug
mode.
These benches are limited by drive performance on a lot of machines, but
still useful as a local tool for iterating on CPU/memory improvements
around this code path.
Anecdotal measurements on Hetzner AX102 (Ryzen 7950xd):
```
ingest-small-values/ingest 128MB/100b seq
time: [1.1160 s 1.1230 s 1.1289 s]
thrpt: [113.38 MiB/s 113.98 MiB/s 114.70 MiB/s]
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) low mild
Benchmarking ingest-small-values/ingest 128MB/100b rand: Warming up for 3.0000 s
Warning: Unable to complete 10 samples in 10.0s. You may wish to increase target time to 18.9s.
ingest-small-values/ingest 128MB/100b rand
time: [1.9001 s 1.9056 s 1.9110 s]
thrpt: [66.982 MiB/s 67.171 MiB/s 67.365 MiB/s]
Benchmarking ingest-small-values/ingest 128MB/100b rand-1024keys: Warming up for 3.0000 s
Warning: Unable to complete 10 samples in 10.0s. You may wish to increase target time to 11.0s.
ingest-small-values/ingest 128MB/100b rand-1024keys
time: [1.0715 s 1.0828 s 1.0937 s]
thrpt: [117.04 MiB/s 118.21 MiB/s 119.46 MiB/s]
ingest-small-values/ingest 128MB/100b seq, no delta
time: [425.49 ms 429.07 ms 432.04 ms]
thrpt: [296.27 MiB/s 298.32 MiB/s 300.83 MiB/s]
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) low mild
ingest-big-values/ingest 128MB/8k seq
time: [373.03 ms 375.84 ms 379.17 ms]
thrpt: [337.58 MiB/s 340.57 MiB/s 343.13 MiB/s]
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) high mild
ingest-big-values/ingest 128MB/8k seq, no delta
time: [81.534 ms 82.811 ms 83.364 ms]
thrpt: [1.4994 GiB/s 1.5095 GiB/s 1.5331 GiB/s]
Found 1 outliers among 10 measurements (10.00%)
```
## 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.
```
While investigating Pageserver logs from the cases where systemd hangs
during shutdown (https://github.com/neondatabase/cloud/issues/11387), I
noticed that even if Pageserver shuts down cleanly[^1], there are
lingering walredo processes.
[^1]: Meaning, pageserver finishes its shutdown procedure and calls
`exit(0)` on its own terms, instead of hitting the systemd unit's
`TimeoutSec=` limit and getting SIGKILLed.
While systemd should never lock up like it does, maybe we can avoid
hitting that bug by cleaning up properly.
Changes
-------
This PR adds a shutdown method to `WalRedoManager` and hooks it up to
tenant shutdown.
We keep track of intent to shutdown through the new `enum
ProcessOnceCell` stored inside the pre-existing `redo_process` field.
A gate is added to keep track of running processes, using the new type
`struct Process`.
Future Work
-----------
Requests that don't need the redo process will not observe the shutdown
(see doc comment).
Doing so would be nice for completeness sake, but doesn't provide much
benefit because `Tenant` and `Timeline` already shut down all walredo
users.
Testing
-------
I did manual testing to confirm that the problem exists before this PR
and that it's gone after.
Setup:
* `neon_local` with a single tenant, create some data using `pgbench`
* ensure walredo process is running, not pid
* watch `strace -e kill,wait4 -f -p "$(pgrep pageserver)"`
* `neon_local pageserver stop`
With this PR, we always observe
```
$ strace -e kill,wait4 -f -p "$(pgrep pageserver)"
...
[pid 591120] --- SIGTERM {si_signo=SIGTERM, si_code=SI_USER, si_pid=591215, si_uid=1000} ---
[pid 591134] kill(591174, SIGKILL) = 0
[pid 591134] wait4(591174, <unfinished ...>
[pid 591142] --- SIGCHLD {si_signo=SIGCHLD, si_code=CLD_KILLED, si_pid=591174, si_uid=1000, si_status=SIGKILL, si_utime=0, si_stime=0} ---
[pid 591134] <... wait4 resumed>[{WIFSIGNALED(s) && WTERMSIG(s) == SIGKILL}], 0, NULL) = 591174
...
+++ exited with 0 +++
```
Before this PR, we'd usually observe just
```
...
[pid 596239] --- SIGTERM {si_signo=SIGTERM, si_code=SI_USER, si_pid=596455, si_uid=1000} ---
...
+++ exited with 0 +++
```
Refs
----
refs https://github.com/neondatabase/cloud/issues/11387
refs https://github.com/neondatabase/neon/issues/7753
This PR is step (1) of removing sync walredo from Pageserver.
Changes:
* Remove the sync impl
* If sync is configured, warn! and use async instead
* Remove the metric that exposes `kind`
* Remove the tenant status API that exposes `kind`
Future Work
-----------
After we've released this change to prod and are sure we won't
roll back, we will
1. update the prod Ansible to remove the config flag from the prod
pageserver.toml.
2. remove the remaining `kind` code in pageserver
These two changes need no release inbetween.
See https://github.com/neondatabase/neon/issues/7753 for details.
- Rename "filename" types which no longer map directly to a filename
(LayerFileName -> LayerName)
- Add a -v1- part to local layer paths to smooth the path to future
updates (we anticipate a -v2- that uses checksums later)
- Rename methods that refer to the string-ized version of a LayerName to
no longer be called "filename"
- Refactor reconcile() function to use a LocalLayerFileMetadata type
that includes the local path, rather than carrying local path separately
in a tuple and unwrap()'ing it later.
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
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.
See the updated `bench_walredo.rs` module comment.
tl;dr: we measure avg latency of single redo operations issues against a
single redo manager from N tokio tasks.
part of https://github.com/neondatabase/neon/issues/6628
(includes two preparatory commits from
https://github.com/neondatabase/neon/pull/5960)
## Problem
To accommodate multiple shards in the same tenant on the same
pageserver, we must include the full TenantShardId in local paths. That
means that all code touching local storage needs to see the
TenantShardId.
## Summary of changes
- Replace `tenant_id: TenantId` with `tenant_shard_id: TenantShardId` on
Tenant, Timeline and RemoteTimelineClient.
- Use TenantShardId in helpers for building local paths.
- Update all the relevant call sites.
This doesn't update absolutely everything: things like PageCache,
TaskMgr, WalRedo are still shard-naive. The purpose of this PR is to
update the core types so that others code can be added/updated
incrementally without churning the most central shared types.
For 2 weeks we've seen rare, spurious, not-reproducible page
reconstruction
failures with PG16 in prod.
One of the commits we deployed this week was
Commit
commit fc467941f9
Author: Joonas Koivunen <joonas@neon.tech>
Date: Wed Oct 4 16:19:19 2023 +0300
walredo: log retryed error (#546)
With the logs from that commit, we learned that some read() or write()
system call that walredo does fails with `EAGAIN`, aka
`Resource temporarily unavailable (os error 11)`.
But we have no idea where exactly in the code we get back that error.
So, use anyhow instead of fake std::io::Error's as an easy way to get
a backtrace when the error happens, and change the logging to print
that backtrace (i.e., use `{:?}` instead of
`utils::error::report_compact_sources(e)`).
The `WalRedoError` type had to go because we add additional `.context()`
further up the call chain before we `{:?}`-print it. That additional
`.context()` further up doesn't see that there's already an
anyhow::Error
inside the `WalRedoError::ApplyWalRecords` variant, and hence captures
another backtrace and prints that one on `{:?}`-print instead of the
original one inside `WalRedoError::ApplyWalRecords`.
If we ever switch back to `report_compact_sources`, we should make sure
we have some other way to uniquely identify the places where we return
an error in the error message.
Fixes#4689 by replacing all of `std::Path` , `std::PathBuf` with
`camino::Utf8Path`, `camino::Utf8PathBuf` in
- pageserver
- safekeeper
- control_plane
- libs/remote_storage
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
Accidentially giving is_incremental=true for ImageLayers costs a lot of
debugging time. Removes all API which would allow to do that. They can
easily be restored later *when needed*.
Split off from #4938.
## Problem
part of https://github.com/neondatabase/neon/pull/4340
## Summary of changes
Remove LayerDescriptor and remove `todo!`. At the same time, this PR
adds `AsLayerDesc` trait for all persistent layers and changed
`LayerFileManager` to have a generic type. For tests, we are now using
`LayerObject`, which is a wrapper around `PersistentLayerDesc`.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
part of https://github.com/neondatabase/neon/issues/4392, continuation
of https://github.com/neondatabase/neon/pull/4408
## Summary of changes
This PR removes all layer objects from LayerMap and moves it to the
timeline struct. In timeline struct, LayerFileManager maps a layer
descriptor to a layer object, and it is stored in the same RwLock as
LayerMap to avoid behavior difference.
Key changes:
* LayerMap now does not have generic, and only stores descriptors.
* In Timeline, we add a new struct called layer mapping.
* Currently, layer mapping is stored in the same lock with layer map.
Every time we retrieve data from the layer map, we will need to map the
descriptor to the actual object.
* Replace_historic is moved to layer mapping's replace, and the return
value behavior is different from before. I'm a little bit unsure about
this part and it would be good to have some comments on that.
* Some test cases are rewritten to adapt to the new interface, and we
can decide whether to remove it in the future because it does not make
much sense now.
* LayerDescriptor is moved to `tests` module and should only be intended
for unit testing / benchmarks.
* Because we now have a usage pattern like "take the guard of lock, then
get the reference of two fields", we want to avoid dropping the
incorrect object when we intend to unlock the lock guard. Therefore, a
new set of helper function `drop_r/wlock` is added. This can be removed
in the future when we finish the refactor.
TODOs after this PR: fully remove RemoteLayer, and move LayerMapping to
a separate LayerCache.
all refactor PRs:
```
#4437 --- #4479 ------------ #4510 (refactor done at this point)
\-- #4455 -- #4502 --/
```
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
part of https://github.com/neondatabase/neon/issues/4392
## Summary of changes
This PR adds a new HashMap that maps persistent layer desc to the layer
object *inside* LayerMap. Originally I directly went towards adding such
layer cache in Timeline, but the changes are too many and cannot be
reviewed as a reasonably-sized PR. Therefore, we take this intermediate
step to change part of the codebase to use persistent layer desc, and
come up with other PRs to move this hash map of layer desc to the
timeline struct.
Also, file_size is now part of the layer desc.
---------
Signed-off-by: Alex Chi <iskyzh@gmail.com>
Co-authored-by: bojanserafimov <bojan.serafimov7@gmail.com>
This reverts commit 732acc5.
Reverted PR: #3869
As noted in PR #4094, we do in fact try to insert duplicates to the
layer map, if L0->L1 compaction is interrupted. We do not have a proper
fix for that right now, and we are in a hurry to make a release to
production, so revert the changes related to this to the state that we
have in production currently. We know that we have a bug here, but
better to live with the bug that we've had in production for a long
time, than rush a fix to production without testing it in staging first.
Cc: #4094, #4088
