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 #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>
refs https://github.com/neondatabase/neon/issues/6989
Problem
-------
After unclean shutdown, we get restarted, start reading the local
filesystem,
and make decisions based on those reads. However, some of the data might
have
not yet been fsynced when the unclean shutdown completed.
Durability matters even though Pageservers are conceptually just a cache
of state in S3. For example:
- the cloud control plane is no control loop => pageserver responses
to tenant attachmentm, etc, needs to be durable.
- the storage controller does not rely on this (as much?)
- we don't have layer file checksumming, so, downloaded+renamed but not
fsynced layer files are technically not to be trusted
- https://github.com/neondatabase/neon/issues/2683
Solution
--------
`syncfs` the tenants directory during startup, before we start reading
from it.
This is a bit overkill because we do remove some temp files
(InMemoryLayer!)
later during startup. Further, these temp files are particularly likely
to
be dirty in the kernel page cache. However, we don't want to refactor
that
cleanup code right now, and the dirty data on pageservers is generally
not that high. Last, with [direct
IO](https://github.com/neondatabase/neon/issues/8130) we're going to
have near-zero kernel page cache anyway quite soon.
## Problem
Pageserver exposes some vectored get related configs which are not in
use.
## Summary of changes
Remove the following pageserver configs: get_impl, get_vectored_impl,
and `validate_get_vectored`.
They are not used in the pageserver since
https://github.com/neondatabase/neon/pull/8601.
Manual overrides have been removed from the aws repo in
https://github.com/neondatabase/aws/pull/1664.
Part of #8130, [RFC: Direct IO For Pageserver](https://github.com/neondatabase/neon/blob/problame/direct-io-rfc/docs/rfcs/034-direct-io-for-pageserver.md)
## Description
Add pageserver config for evaluating/enabling direct I/O.
- Disabled: current default, uses buffered io as is.
- Evaluate: still uses buffered io, but could do alignment checking and
perf simulation (pad latency by direct io RW to a fake file).
- Enabled: uses direct io, behavior on alignment error is configurable.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
Since the introduction of sharding, the protocol handling loop in
`handle_pagerequests` cannot know anymore which concrete
`Tenant`/`Timeline` object any of the incoming `PagestreamFeMessage`
resolves to.
In fact, one message might resolve to one `Tenant`/`Timeline` while
the next one may resolve to another one.
To avoid going to tenant manager, we added the `shard_timelines` which
acted as an ever-growing cache that held timeline gate guards open for
the lifetime of the connection.
The consequence of holding the gate guards open was that we had to be
sensitive to every cached `Timeline::cancel` on each interaction with
the network connection, so that Timeline shutdown would not have to wait
for network connection interaction.
We can do better than that, meaning more efficiency & better
abstraction.
I proposed a sketch for it in
* https://github.com/neondatabase/neon/pull/8286
and this PR implements an evolution of that sketch.
The main idea is is that `mod page_service` shall be solely concerned
with the following:
1. receiving requests by speaking the protocol / pagestream subprotocol
2. dispatching the request to a corresponding method on the correct
shard/`Timeline` object
3. sending response by speaking the protocol / pagestream subprotocol.
The cancellation sensitivity responsibilities are clear cut:
* while in `page_service` code, sensitivity to page_service cancellation
is sufficient
* while in `Timeline` code, sensitivity to `Timeline::cancel` is
sufficient
To enforce these responsibilities, we introduce the notion of a
`timeline::handle::Handle` to a `Timeline` object that is checked out
from a `timeline::handle::Cache` for **each request**.
The `Handle` derefs to `Timeline` and is supposed to be used for a
single async method invocation on `Timeline`.
See the lengthy doc comment in `mod handle` for details of the design.
part of https://github.com/neondatabase/neon/issues/8184
# Problem
We want to bypass PS PageCache for all data block reads, but
`compact_level0_phase1` currently uses `ValueRef::load` to load the WAL
records from delta layers.
Internally, that maps to `FileBlockReader:read_blk` which hits the
PageCache
[here](e78341e1c2/pageserver/src/tenant/block_io.rs (L229-L236)).
# Solution
This PR adds a mode for `compact_level0_phase1` that uses the
`MergeIterator` for reading the `Value`s from the delta layer files.
`MergeIterator` is a streaming k-merge that uses vectored blob_io under
the hood, which bypasses the PS PageCache for data blocks.
Other notable changes:
* change the `DiskBtreeReader::into_stream` to buffer the node, instead
of holding a `PageCache` `PageReadGuard`.
* Without this, we run out of page cache slots in
`test_pageserver_compaction_smoke`.
* Generally, `PageReadGuard`s aren't supposed to be held across await
points, so, this is a general bugfix.
# Testing / Validation / Performance
`MergeIterator` has not yet been used in production; it's being
developed as part of
* https://github.com/neondatabase/neon/issues/8002
Therefore, this PR adds a validation mode that compares the existing
approach's value iterator with the new approach's stream output, item by
item.
If they're not identical, we log a warning / fail the unit/regression
test.
To avoid flooding the logs, we apply a global rate limit of once per 10
seconds.
In any case, we use the existing approach's value.
Expected performance impact that will be monitored in staging / nightly
benchmarks / eventually pre-prod:
* with validation:
* increased CPU usage
* ~doubled VirtualFile read bytes/second metric
* no change in disk IO usage because the kernel page cache will likely
have the pages buffered on the second read
* without validation:
* slightly higher DRAM usage because each iterator participating in the
k-merge has a dedicated buffer (as opposed to before, where compactions
would rely on the PS PageCaceh as a shared evicting buffer)
* less disk IO if previously there were repeat PageCache misses (likely
case on a busy production Pageserver)
* lower CPU usage: PageCache out of the picture, fewer syscalls are made
(vectored blob io batches reads)
# Rollout
The new code is used with validation mode enabled-by-default.
This gets us validation everywhere by default, specifically in
- Rust unit tests
- Python tests
- Nightly pagebench (shouldn't really matter)
- Staging
Before the next release, I'll merge the following aws.git PR that
configures prod to continue using the existing behavior:
* https://github.com/neondatabase/aws/pull/1663
# Interactions With Other Features
This work & rollout should complete before Direct IO is enabled because
Direct IO would double the IOPS & latency for each compaction read
(#8240).
# Future Work
The streaming k-merge's memory usage is proportional to the amount of
memory per participating layer.
But `compact_level0_phase1` still loads all keys into memory for
`all_keys_iter`.
Thus, it continues to have active memory usage proportional to the
number of keys involved in the compaction.
Future work should replace `all_keys_iter` with a streaming keys
iterator.
This PR has a draft in its first commit, which I later reverted because
it's not necessary to achieve the goal of this PR / issue #8184.
## Problem
In `test_basebackup_with_high_slru_count`, the pageserver is sometimes
mysteriously hanging on startup, having been started+stopped earlier in
the test setup while populating template tenant data.
- #7586
We can't see why this is hanging in this particular test. The test does
some weird stuff though, like attaching a load of broken tenants and
then doing a SIGQUIT kill of a pageserver.
## Summary of changes
- Attach tenants normally instead of doing a failpoint dance to attach
them as broken
- Shut the pageserver down gracefully during init instead of using
immediate mode
- Remove the "sequential" variant of the unstable test, as this is going
away soon anyway
- Log before trying to acquire lock file, so that if it hangs we have a
clearer sense of if that's really where it's hanging. It seems like it
is, but that code does a non-blocking flock so it's surprising.
## Motivation & Context
We want to move away from `task_mgr` towards explicit tracking of child
tasks.
This PR is extracted from https://github.com/neondatabase/neon/pull/8339
where I refactor `PageRequestHandler` to not depend on task_mgr anymore.
## Changes
This PR refactors all global tasks but `PageRequestHandler` to use some
combination of `JoinHandle`/`JoinSet` + `CancellationToken`.
The `task_mgr::spawn(.., shutdown_process_on_error)` functionality is
preserved through the new `exit_on_panic_or_error` wrapper.
Some global tasks were not using it before, but as of this PR, they are.
The rationale is that all global tasks are relevant for correct
operation of the overall Neon system in one way or another.
## Future Work
After #8339, we can make `task_mgr::spawn` require a `TenantId` instead
of an `Option<TenantId>` which concludes this step of cleanup work and
will help discourage future usage of task_mgr for global tasks.
PR #8299 has switched the storage scrubber to use
`DefaultCredentialsChain`. Now we do this for `remote_storage`, as it
allows us to use `remote_storage` from inside kubernetes. Most of the
diff is due to `GenericRemoteStorage::from_config` becoming `async fn`.
We're removing the usage of this long-meaningless config field in
https://github.com/neondatabase/aws/pull/1599
Once that PR has been deployed to staging and prod, we can merge this
PR.
I want to fix bugs in `page_service`
([issue](https://github.com/neondatabase/neon/issues/7427)) and the
`import basebackup` / `import wal` stand in the way / make the
refactoring more complicated.
We don't use these methods anyway in practice, but, there have been some
objections to removing the functionality completely.
So, this PR preserves the existing functionality but moves it into the
HTTP management API.
Note that I don't try to fix existing bugs in the code, specifically not
fixing
* it only ever worked correctly for unsharded tenants
* it doesn't clean up on error
All errors are mapped to `ApiError::InternalServerError`.
## Problem
- Resident memory on long running pageserver processes tends to climb:
memory fragmentation is suspected.
- Total resident memory may be a limiting factor for running on smaller
nodes.
## Summary of changes
- As a low-energy experiment, switch the pageserver to use jemalloc (not
a net-new dependency, proxy already use it)
- Decide at end of week whether to revert before next release.
part of https://github.com/neondatabase/neon/issues/7418
# Motivation
(reproducing #7418)
When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.
In benchmarks (https://github.com/neondatabase/neon/pull/7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.
# High-Level Changes
Add a new mode for L0 flush that works as follows:
* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.
# Implementation Details
The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
https://github.com/neondatabase/neon/issues/8183
# Correctness
The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.
The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
https://github.com/neondatabase/neon/pull/8186 to make it less
error-prone, but still, careful review is requested.
# Performance
I manually measured single-client ingest performance from `pgbench -i
...`.
Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4
tl;dr:
* no speed improvements during ingest, but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
* (that's why I'm working on this)
* noticable but modestly lower CPU time
This is good enough for merging this PR because the changes require
opt-in.
We'll do more testing in staging & pre-prod.
# Stability / Monitoring
**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](23827c6b0d/pageserver/src/tenant/timeline.rs (L5741-L5743))
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.
**metrics**: we do have the
[flush_time_histo](23827c6b0d/pageserver/src/tenant/timeline.rs (L3725-L3726)),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
## Problem
This is tech debt from when shard splitting was implemented, to handle
more nicely the edge case of a client reconnect at the moment of the
split.
During shard splits, there were edge cases where we could incorrectly
return NotFound to a getpage@lsn request, prompting an unwanted
reconnect/backoff from the client.
It is already the case that parent shards during splits are marked
InProgress before child shards are created, so `resolve_attached_shard`
will not match on them, thereby implicitly preferring child shards
(good).
However, we were not doing any elegant handling of InProgress in
general: `get_active_tenant_with_timeout` was previously mostly dead
code: it was inspecting the slot found by `resolve_attached_shard` and
maybe waiting for InProgress, but that path is never taken because since
ef7c9c2ccc the resolve function only ever
returns attached slots.
Closes: https://github.com/neondatabase/neon/issues/7044
## Summary of changes
- Change return value of `resolve_attached_shard` to distinguish between
true NotFound case, and the case where we skipped slots that were
InProgress.
- Rework `get_active_tenant_with_timeout` to loop over calling
resolve_attached_shard, waiting if it sees an InProgress result.
The resulting behavior during a shard split is:
- If we look up a shard early in split when parent is InProgress but
children aren't created yet, we'll wait for the parent to be shut down.
This corresponds to the part of the split where we wait for LSNs to
catch up: so a small delay to the request, but a clean enough handling.
- If we look up a shard while child shards are already present, we will
match on those shards rather than the parent, as intended.
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.
## Problem
This is historical baggage from when the pageserver could be run with
local disk only: we had a bunch of places where we had to treat remote
storage as optional.
Closes: https://github.com/neondatabase/neon/issues/6890
## Changes
- Remove Option<> around remote storage (in
https://github.com/neondatabase/neon/pull/7722 we made remote storage
clearly mandatory)
- Remove code for deleting old metadata files: they're all gone now.
- Remove other references to metadata files when loading directories, as
none exist.
I checked last 14 days of logs for "found legacy metadata", there are no
instances.
## Problem
Since https://github.com/neondatabase/neon/pull/6769, the pageserver is
intentionally not usable without remote storage: it's purpose is to act
as a cache to an object store, rather than as a source of truth in its
own right.
## Summary of changes
- Make remote storage configuration mandatory: the pageserver will
refuse to start if it is not provided.
This is a precursor that will make it safe to subsequently remove all
the internal Option<>s
## Problem
We are currently supporting two read paths. No bueno.
## Summary of changes
High level: use vectored read path to serve get page requests - gated by
`get_impl` config
Low level:
1. Add ps config, `get_impl` to specify which read path to use when
serving get page requests
2. Fix base cached image handling for the vectored read path. This was
subtly broken: previously we
would not mark keys that went past their cached lsn as complete. This is
a self standing change which
could be its own PR, but I've included it here because writing separate
tests for it is tricky.
3. Fork get page to use either the legacy or vectored implementation
4. Validate the use of vectored read path when serving get page requests
against the legacy implementation.
Controlled by `validate_vectored_get` ps config.
5. Use the vectored read path to serve get page requests in tests (with
validation).
## Note
Since the vectored read path does not go through the page cache to read
buffers, this change also amounts to a removal of the buffer page cache. Materialized page cache
is still used.
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
It's just unnecessary to use spawn_blocking there, and with
https://github.com/neondatabase/neon/pull/7331 , it will result in
really just one executor thread when enabling one-runtime with
current_thread executor.
## Problem
This is a refactor.
This PR was a precursor to a much smaller change
e5bd602dc1,
where as I was writing it I found that we were not far from getting rid
of the last non-deprecated code paths that use `mgr::` scoped functions
to get at the TenantManager state.
We're almost done cleaning this up as per
https://github.com/neondatabase/neon/issues/5796. The only significant
remaining mgr:: item is `get_active_tenant_with_timeout`, which is
page_service's path for fetching tenants.
## Summary of changes
- Remove the bool argument to get_attached_tenant_shard: this was almost
always false from API use cases, and in cases when it was true, it was
readily replacable with an explicit check of the returned tenant's
status.
- Rather than letting the timeline eviction task query any tenant it
likes via `mgr::`, pass an `Arc<Tenant>` into the task. This is still an
ugly circular reference, but should eventually go away: either when we
switch to exclusively using disk usage eviction, or when we change
metadata storage to avoid the need to imitate layer accesses.
- Convert all the mgr::get_tenant call sites to use
TenantManager::get_attached_tenant_shard
- Move list_tenants into TenantManager.
Before this PR, each core had 3 executor threads from 3 different
runtimes. With this PR, we just have one runtime, with one thread per
core. Switching to a single tokio runtime should reduce that effective
over-commit of CPU and in theory help with tail latencies -- iff all
tokio tasks are well-behaved and yield to the runtime regularly.
Are All Tasks Well-Behaved? Are We Ready?
-----------------------------------------
Sadly there doesn't seem to be good out-of-the box tokio tooling to
answer this question.
We *believe* all tasks are well behaved in today's code base, as of the
switch to `virtual_file_io_engine = "tokio-epoll-uring"` in production
(https://github.com/neondatabase/aws/pull/1121).
The only remaining executor-thread-blocking code is walredo and some
filesystem namespace operations.
Filesystem namespace operations work is being tracked in #6663 and not
considered likely to actually block at this time.
Regarding walredo, it currently does a blocking `poll` for read/write to
the pipe file descriptors we use for IPC with the walredo process.
There is an ongoing experiment to make walredo async (#6628), but it
needs more time because there are surprisingly tricky trade-offs that
are articulated in that PR's description (which itself is still WIP).
What's relevant for *this* PR is that
1. walredo is always CPU-bound
2. production tail latencies for walredo request-response
(`pageserver_wal_redo_seconds_bucket`) are
- p90: with few exceptions, low hundreds of micro-seconds
- p95: except on very packed pageservers, below 1ms
- p99: all below 50ms, vast majority below 1ms
- p99.9: almost all around 50ms, rarely at >= 70ms
- [Dashboard
Link](https://neonprod.grafana.net/d/edgggcrmki3uof/2024-03-walredo-latency?orgId=1&var-ds=ZNX49CDVz&var-pXX_by_instance=0.9&var-pXX_by_instance=0.99&var-pXX_by_instance=0.95&var-adhoc=instance%7C%21%3D%7Cpageserver-30.us-west-2.aws.neon.tech&var-per_instance_pXX_max_seconds=0.0005&from=1711049688777&to=1711136088777)
The ones below 1ms are below our current threshold for when we start
thinking about yielding to the executor.
The tens of milliseconds stalls aren't great, but, not least because of
the implicit overcommit of CPU by the three runtimes, we can't be sure
whether these tens of milliseconds are inherently necessary to do the
walredo work or whether we could be faster if there was less contention
for CPU.
On the first item (walredo being always CPU-bound work): it means that
walredo processes will always compete with the executor threads.
We could yield, using async walredo, but then we hit the trade-offs
explained in that PR.
tl;dr: the risk of stalling executor threads through blocking walredo
seems low, and switching to one runtime cleans up one potential source
for higher-than-necessary stall times (explained in the previous
paragraphs).
Code Changes
------------
- Remove the 3 different runtime definitions.
- Add a new definition called `THE_RUNTIME`.
- Use it in all places that previously used one of the 3 removed
runtimes.
- Remove the argument from `task_mgr`.
- Fix failpoint usage where `pausable_failpoint!` should have been used.
We encountered some actual failures because of this, e.g., hung
`get_metric()` calls during test teardown that would client-timeout
after 300s.
As indicated by the comment above `THE_RUNTIME`, we could take this
clean-up further.
But before we create so much churn, let's first validate that there's no
perf regression.
Performance
-----------
We will test this in staging using the various nightly benchmark runs.
However, the worst-case impact of this change is likely compaction
(=>image layer creation) competing with compute requests.
Image layer creation work can't be easily generated & repeated quickly
by pagebench.
So, we'll simply watch getpage & basebackup tail latencies in staging.
Additionally, I have done manual benchmarking using pagebench.
Report:
https://neondatabase.notion.site/2024-03-23-oneruntime-change-benchmarking-22a399c411e24399a73311115fb703ec?pvs=4
Tail latencies and throughput are marginally better (no regression =
good).
Except in a workload with 128 clients against one tenant.
There, the p99.9 and p99.99 getpage latency is about 2x worse (at
slightly lower throughput).
A dip in throughput every 20s (compaction_period_ is clearly visible,
and probably responsible for that worse tail latency.
This has potential to improve with async walredo, and is an edge case
workload anyway.
Future Work
-----------
1. Once this change has shown satisfying results in production, change
the codebase to use the ambient runtime instead of explicitly
referencing `THE_RUNTIME`.
2. Have a mode where we run with a single-threaded runtime, so we
uncover executor stalls more quickly.
3. Switch or write our own failpoints library that is async-native:
https://github.com/neondatabase/neon/issues/7216
## Problem
The service that receives consumption metrics has lower availability
than S3. Writing metrics to S3 improves their availability.
Closes: https://github.com/neondatabase/cloud/issues/9824
## Summary of changes
- The same data as consumption metrics POST bodies is also compressed
and written to an S3 object with a timestamp-formatted path.
- Set `metric_collection_bucket` (same format as `remote_storage`
config) to configure the location to write to
## Problem
Tenant deletion had a couple of TODOs where we weren't using proper
cancellation tokens that would have aborted the deletions during process
shutdown.
## Summary of changes
- Refactor enough that deletion/shutdown code has access to the
TenantManager's cancellation toke
- Use that cancellation token in tenant deletion instead of dummy
tokens.
fixes https://github.com/neondatabase/neon/issues/7116
Changes:
- refactor PageServerConfigBuilder: support not-set values
- implement runtime feature test
- use runtime feature test to determine `virtual_file_io_engine` if not
explicitly configured in the config
- log the effective engine at startup
- drive-by: improve assertion messages in `test_pageserver_init_node_id`
This needed a tiny bit of tokio-epoll-uring work, hence bumping it.
Changelog:
```
git log --no-decorate --oneline --reverse 868d2c42b5d54ca82fead6e8f2f233b69a540d3e..342ddd197a060a8354e8f11f4d12994419fff939
c7a74c6 Bump mio from 0.8.8 to 0.8.11
4df3466 Bump mio from 0.8.8 to 0.8.11 (#47)
342ddd1 lifecycle: expose `LaunchResult` enum (#49)
```
## Problem
Before this PR, `Timeline::get_vectored` would be throttled twice if the
sequential option was enabled or if validation was enabled.
Also, `pageserver_get_vectored_seconds` included the time spent in the
throttle, which turns out to be undesirable for what we use that metric
for.
## Summary of changes
Double-throttle:
* Add `Timeline::get0` method which is unthrottled.
* Use that method from within the `Timeline::get_vectored` code path.
Metric:
* return throttled time from `throttle()` method
* deduct the value from the observed time
* globally rate-limited logging of duration subtraction errors, like in
all other places that do the throttled-time deduction from observations
This refactoring makes it easier to experimentally replace
BACKGROUND_RUNTIME with a single-threaded runtime. Found this useful
[during benchmarking](https://github.com/neondatabase/neon/pull/6555).
Follows #6123
Closes: https://github.com/neondatabase/neon/issues/5342
The approach here is to avoid using `Layer` from secondary tenants, and
instead make the eviction types (e.g. `EvictionCandidate`) have a
variant that carries a Layer for attached tenants, and a different
variant for secondary tenants.
Other changes:
- EvictionCandidate no longer carries a `Timeline`: this was only used
for providing a witness reference to remote timeline client.
- The types for returning eviction candidates are all in
disk_usage_eviction_task.rs now, whereas some of them were in
timeline.rs before.
- The EvictionCandidate type replaces LocalLayerInfoForDiskUsageEviction
type, which was basically the same thing.
Dependency (commits inline):
https://github.com/neondatabase/neon/pull/5842
## Problem
Secondary mode tenants need a manifest of what to download. Ultimately
this will be some kind of heat-scored set of layers, but as a robust
first step we will simply use the set of resident layers: secondary
tenant locations will aim to match the on-disk content of the attached
location.
## Summary of changes
- Add heatmap types representing the remote structure
- Add hooks to Tenant/Timeline for generating these heatmaps
- Create a new `HeatmapUploader` type that is external to `Tenant`, and
responsible for walking the list of attached tenants and scheduling
heatmap uploads.
Notes to reviewers:
- Putting the logic for uploads (and later, secondary mode downloads)
outside of `Tenant` is an opinionated choice, motivated by:
- Enable future smarter scheduling of operations, e.g. uploading the
stalest tenant first, rather than having all tenants compete for a fair
semaphore on a first-come-first-served basis. Similarly for downloads,
we may wish to schedule the tenants with the hottest un-downloaded
layers first.
- Enable accessing upload-related state without synchronization (it
belongs to HeatmapUploader, rather than being some Mutex<>'d part of
Tenant)
- Avoid further expanding the scope of Tenant/Timeline types, which are
already among the largest in the codebase
- You might reasonably wonder how much of the uploader code could be a
generic job manager thing. Probably some of it: but let's defer pulling
that out until we have at least two users (perhaps secondary downloads
will be the second one) to highlight which bits are really generic.
Compromises:
- Later, instead of using digests of heatmaps to decide whether anything
changed, I would prefer to avoid walking the layers in tenants that
don't have changes: tracking that will be a bit invasive, as it needs
input from both remote_timeline_client and Layer.
## Problem
The cancellation code was confusing and error prone (as seen before in
our memory leaks).
## Summary of changes
* Use the new `TaskTracker` primitve instead of JoinSet to gracefully
wait for tasks to shutdown.
* Updated libs/utils/completion to use `TaskTracker`
* Remove `tokio::select` in favour of `futures::future::select` in a
specialised `run_until_cancelled()` helper function
Problem
-------
Before this PR, there was no concurrency limit on initial logical size
computations.
While logical size computations are lazy in theory, in practice
(production), they happen in a short timeframe after restart.
This means that on a PS with 20k tenants, we'd have up to 20k concurrent
initial logical size calculation requests.
This is self-inflicted needless overload.
This hasn't been a problem so far because the `.await` points on the
logical size calculation path never return `Pending`, hence we have a
natural concurrency limit of the number of executor threads.
But, as soon as we return `Pending` somewhere in the logical size
calculation path, other concurrent tasks get scheduled by tokio.
If these other tasks are also logical size calculations, they eventually
pound on the same bottleneck.
For example, in #5479, we want to switch the VirtualFile descriptor
cache to a `tokio::sync::RwLock`, which makes us return `Pending`, and
without measures like this patch, after PS restart, VirtualFile
descriptor cache thrashes heavily for 2 hours until all the logical size
calculations have been computed and the degree of concurrency /
concurrent VirtualFile operations is down to regular levels.
See the *Experiment* section below for details.
<!-- Experiments (see below) show that plain #5479 causes heavy
thrashing of the VirtualFile descriptor cache.
The high degree of concurrency is too much for
In the case of #5479 the VirtualFile descriptor cache size starts
thrashing heavily.
-->
Background
----------
Before this PR, initial logical size calculation was spawned lazily on
first call to `Timeline::get_current_logical_size()`.
In practice (prod), the lazy calculation is triggered by
`WalReceiverConnectionHandler` if the timeline is active according to
storage broker, or by the first iteration of consumption metrics worker
after restart (`MetricsCollection`).
The spawns by walreceiver are high-priority because logical size is
needed by Safekeepers (via walreceiver `PageserverFeedback`) to enforce
the project logical size limit.
The spawns by metrics collection are not on the user-critical path and
hence low-priority. [^consumption_metrics_slo]
[^consumption_metrics_slo]: We can't delay metrics collection
indefintely because there are TBD internal SLOs tied to metrics
collection happening in a timeline manner
(https://github.com/neondatabase/cloud/issues/7408). But let's ignore
that in this issue.
The ratio of walreceiver-initiated spawns vs
consumption-metrics-initiated spawns can be reconstructed from logs
(`spawning logical size computation from context of task kind {:?}"`).
PR #5995 and #6018 adds metrics for this.
First investigation of the ratio lead to the discovery that walreceiver
spawns 75% of init logical size computations.
That's because of two bugs:
- In Safekeepers: https://github.com/neondatabase/neon/issues/5993
- In interaction between Pageservers and Safekeepers:
https://github.com/neondatabase/neon/issues/5962
The safekeeper bug is likely primarily responsible but we don't have the
data yet. The metrics will hopefully provide some insights.
When assessing production-readiness of this PR, please assume that
neither of these bugs are fixed yet.
Changes In This PR
------------------
With this PR, initial logical size calculation is reworked as follows:
First, all initial logical size calculation task_mgr tasks are started
early, as part of timeline activation, and run a retry loop with long
back-off until success. This removes the lazy computation; it was
needless complexity because in practice, we compute all logical sizes
anyways, because consumption metrics collects it.
Second, within the initial logical size calculation task, each attempt
queues behind the background loop concurrency limiter semaphore. This
fixes the performance issue that we pointed out in the "Problem" section
earlier.
Third, there is a twist to queuing behind the background loop
concurrency limiter semaphore. Logical size is needed by Safekeepers
(via walreceiver `PageserverFeedback`) to enforce the project logical
size limit. However, we currently do open walreceiver connections even
before we have an exact logical size. That's bad, and I'll build on top
of this PR to fix that
(https://github.com/neondatabase/neon/issues/5963). But, for the
purposes of this PR, we don't want to introduce a regression, i.e., we
don't want to provide an exact value later than before this PR. The
solution is to introduce a priority-boosting mechanism
(`GetLogicalSizePriority`), allowing callers of
`Timeline::get_current_logical_size` to specify how urgently they need
an exact value. The effect of specifying high urgency is that the
initial logical size calculation task for the timeline will skip the
concurrency limiting semaphore. This should yield effectively the same
behavior as we had before this PR with lazy spawning.
Last, the priority-boosting mechanism obsoletes the `init_order`'s grace
period for initial logical size calculations. It's a separate commit to
reduce the churn during review. We can drop that commit if people think
it's too much churn, and commit it later once we know this PR here
worked as intended.
Experiment With #5479
---------------------
I validated this PR combined with #5479 to assess whether we're making
forward progress towards asyncification.
The setup is an `i3en.3xlarge` instance with 20k tenants, each with one
timeline that has 9 layers.
All tenants are inactive, i.e., not known to SKs nor storage broker.
This means all initial logical size calculations are spawned by
consumption metrics `MetricsCollection` task kind.
The consumption metrics worker starts requesting logical sizes at low
priority immediately after restart. This is achieved by deleting the
consumption metrics cache file on disk before starting
PS.[^consumption_metrics_cache_file]
[^consumption_metrics_cache_file] Consumption metrics worker persists
its interval across restarts to achieve persistent reporting intervals
across PS restarts; delete the state file on disk to get predictable
(and I believe worst-case in terms of concurrency during PS restart)
behavior.
Before this patch, all of these timelines would all do their initial
logical size calculation in parallel, leading to extreme thrashing in
page cache and virtual file cache.
With this patch, the virtual file cache thrashing is reduced
significantly (from 80k `open`-system-calls/second to ~500
`open`-system-calls/second during loading).
### Critique
The obvious critique with above experiment is that there's no skipping
of the semaphore, i.e., the priority-boosting aspect of this PR is not
exercised.
If even just 1% of our 20k tenants in the setup were active in
SK/storage_broker, then 200 logical size calculations would skip the
limiting semaphore immediately after restart and run concurrently.
Further critique: given the two bugs wrt timeline inactive vs active
state that were mentioned in the Background section, we could have 75%
of our 20k tenants being (falsely) active on restart.
So... (next section)
This Doesn't Make Us Ready For Async VirtualFile
------------------------------------------------
This PR is a step towards asynchronous `VirtualFile`, aka, #5479 or even
#4744.
But it doesn't yet enable us to ship #5479.
The reason is that this PR doesn't limit the amount of high-priority
logical size computations.
If there are many high-priority logical size calculations requested,
we'll fall over like we did if #5479 is applied without this PR.
And currently, at very least due to the bugs mentioned in the Background
section, we run thousands of high-priority logical size calculations on
PS startup in prod.
So, at a minimum, we need to fix these bugs.
Then we can ship #5479 and #4744, and things will likely be fine under
normal operation.
But in high-traffic situations, overload problems will still be more
likely to happen, e.g., VirtualFile cache descriptor thrashing.
The solution candidates for that are orthogonal to this PR though:
* global concurrency limiting
* per-tenant rate limiting => #5899
* load shedding
* scaling bottleneck resources (fd cache size (neondatabase/cloud#8351),
page cache size(neondatabase/cloud#8351), spread load across more PSes,
etc)
Conclusion
----------
Even with the remarks from in the previous section, we should merge this
PR because:
1. it's an improvement over the status quo (esp. if the aforementioned
bugs wrt timeline active / inactive are fixed)
2. it prepares the way for
https://github.com/neondatabase/neon/pull/6010
3. it gets us close to shipping #5479 and #4744
## Problem
#5900
## Summary of changes
Added cancellation token as param in all relevant code paths and actually used it in the find_lsn_for_timestamp main loop
(part of the getpage benchmarking epic #5771)
The plan is to make the benchmarking tool log on stderr and emit results
as JSON on stdout. That way, the test suite can simply take captures
stdout and json.loads() it, while interactive users of the benchmarking
tool have a reasonable experience as well.
Existing logging users continue to print to stdout, so, this change
should be a no-op functionally and performance-wise.
## Problem
For quickly rotating JWT secrets, we want to be able to reload the JWT
public key file in the pageserver, and also support multiple JWT keys.
See #4897.
## Summary of changes
* Allow directories for the `auth_validation_public_key_path` config
param instead of just files. for the safekeepers, all of their config options
also support multiple JWT keys.
* For the pageservers, make the JWT public keys easily globally swappable
by using the `arc-swap` crate.
* Add an endpoint to the pageserver, triggered by a POST to
`/v1/reload_auth_validation_keys`, that reloads the JWT public keys from
the pre-configured path (for security reasons, you cannot upload any
keys yourself).
Fixes#4897
---------
Co-authored-by: Heikki Linnakangas <heikki@neon.tech>
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
## Problem
See: https://github.com/neondatabase/neon/issues/5796
## Summary of changes
Completing the refactor is quite verbose and can be done in stages: each
interface that is currently called directly from a top-level mgr.rs
function can be moved into TenantManager once the relevant subsystems
have access to it.
Landing the initial change to create of TenantManager is useful because
it enables new code to use it without having to be altered later, and
sets us up to incrementally fix the existing code to use an explicit
Arc<TenantManager> instead of relying on the static TENANTS.
- Add a new util `project_build_tag` macro, similar to
`project_git_version`
- Update the `set_build_info_metric` to accept and make use of
`build_tag` info
- Update all codes which use the `set_build_info_metric`
## Problem
Loading tenants shouldn't hang. However, if it does, we shouldn't let
one hung tenant prevent the entire process from starting background
jobs.
## Summary of changes
Generalize the timeout mechanism that we already applied to loading
initial logical sizes: each phase in startup where we wait for a barrier
is subject to a timeout, and startup will proceed if it doesn't complete
within timeout.
Startup metrics will still reflect the time when a phase actually
completed, rather than when we skipped it.
The code isn't the most beautiful, but that kind of reflects the
awkwardness of await'ing on a future and then stashing it to await again
later if we time out. I could imagine making this cleaner in future by
waiting on a structure that doesn't self-destruct on wait() the way
Barrier does, then make InitializationOrder into a structure that
manages the series of waits etc.
## Problem
- QueryError always logged at error severity, even though disconnections
are not true errors.
- QueryError type is not expressive enough to distinguish actual errors
from shutdowns.
- In some functions we're returning Ok(()) on shutdown, in others we're
returning an error
## Summary of changes
- Add QueryError::Shutdown and use it in places we check for
cancellation
- Adopt consistent Result behavior: disconnects and shutdowns are always
QueryError, not ok
- Transform shutdown+disconnect errors to Ok(()) at the very top of the
task that runs query handler
- Use the postgres protocol error code for "admin shutdown" in responses
to clients when we are shutting down.
Closes: #5517
## Problem
Currently it's unclear how much of the `initial_tenant_load` period is
in S3 objects, and therefore how impactful it is to make changes to
remote operations during startup.
## Summary of changes
- `Tenant::load` is refactored to load remote indices in parallel and to
wait for all these remote downloads to finish before it proceeds to
construct any `Timeline` objects.
- `pageserver_startup_duration_seconds` gets a new `phase` value of
`initial_tenant_load_remote` which counts the time from startup to when
the last tenant finishes loading remote content.
- `test_pageserver_restart` is extended to validate this phase. The
previous version of the test was relying on order of dict entries, which
stopped working when adding a phase, so this is refactored a bit.
- `test_pageserver_restart` used to explicitly create a branch, now it
uses the default initial_timeline. This avoids startup getting held up
waiting for logical sizes, when one of the branches is not in use.
