Before this PR, the `nix::poll::poll` call would stall the executor.
This PR refactors the `walredo::process` module to allow for different
implementations, and adds a new `async` implementation which uses
`tokio::process::ChildStd{in,out}` for IPC.
The `sync` variant remains the default for now; we'll do more testing in
staging and gradual rollout to prod using the config variable.
Performance
-----------
I updated `bench_walredo.rs`, demonstrating that a single `async`-based
walredo manager used by N=1...128 tokio tasks has lower latency and
higher throughput.
I further did manual less-micro-benchmarking in the real pageserver
binary.
Methodology & results are published here:
https://neondatabase.notion.site/2024-04-08-async-walredo-benchmarking-8c0ed3cc8d364a44937c4cb50b6d7019?pvs=4
tl;dr:
- use pagebench against a pageserver patched to answer getpage request &
small-enough working set to fit into PS PageCache / kernel page cache.
- compare knee in the latency/throughput curve
- N tenants, each 1 pagebench clients
- sync better throughput at N < 30, async better at higher N
- async generally noticable but not much worse p99.X tail latencies
- eyeballing CPU efficiency in htop, `async` seems significantly more
CPU efficient at ca N=[0.5*ncpus, 1.5*ncpus], worse than `sync` outside
of that band
Mental Model For Walredo & Scheduler Interactions
-------------------------------------------------
Walredo is CPU-/DRAM-only work.
This means that as soon as the Pageserver writes to the pipe, the
walredo process becomes runnable.
To the Linux kernel scheduler, the `$ncpus` executor threads and the
walredo process thread are just `struct task_struct`, and it will divide
CPU time fairly among them.
In `sync` mode, there are always `$ncpus` runnable `struct task_struct`
because the executor thread blocks while `walredo` runs, and the
executor thread becomes runnable when the `walredo` process is done
handling the request.
In `async` mode, the executor threads remain runnable unless there are
no more runnable tokio tasks, which is unlikely in a production
pageserver.
The above means that in `sync` mode, there is an implicit concurrency
limit on concurrent walredo requests (`$num_runtimes *
$num_executor_threads_per_runtime`).
And executor threads do not compete in the Linux kernel scheduler for
CPU time, due to the blocked-runnable-ping-pong.
In `async` mode, there is no concurrency limit, and the walredo tasks
compete with the executor threads for CPU time in the kernel scheduler.
If we're not CPU-bound, `async` has a pipelining and hence throughput
advantage over `sync` because one executor thread can continue
processing requests while a walredo request is in flight.
If we're CPU-bound, under a fair CPU scheduler, the *fixed* number of
executor threads has to share CPU time with the aggregate of walredo
processes.
It's trivial to reason about this in `sync` mode due to the
blocked-runnable-ping-pong.
In `async` mode, at 100% CPU, the system arrives at some (potentially
sub-optiomal) equilibrium where the executor threads get just enough CPU
time to fill up the remaining CPU time with runnable walredo process.
Why `async` mode Doesn't Limit Walredo Concurrency
--------------------------------------------------
To control that equilibrium in `async` mode, one may add a tokio
semaphore to limit the number of in-flight walredo requests.
However, the placement of such a semaphore is non-trivial because it
means that tasks queuing up behind it hold on to their request-scoped
allocations.
In the case of walredo, that might be the entire reconstruct data.
We don't limit the number of total inflight Timeline::get (we only
throttle admission).
So, that queue might lead to an OOM.
The alternative is to acquire the semaphore permit *before* collecting
reconstruct data.
However, what if we need to on-demand download?
A combination of semaphores might help: one for reconstruct data, one
for walredo.
The reconstruct data semaphore permit is dropped after acquiring the
walredo semaphore permit.
This scheme effectively enables both a limit on in-flight reconstruct
data and walredo concurrency.
However, sizing the amount of permits for the semaphores is tricky:
- Reconstruct data retrieval is a mix of disk IO and CPU work.
- If we need to do on-demand downloads, it's network IO + disk IO + CPU
work.
- At this time, we have no good data on how the wall clock time is
distributed.
It turns out that, in my benchmarking, the system worked fine without a
semaphore. So, we're shipping async walredo without one for now.
Future Work
-----------
We will do more testing of `async` mode and gradual rollout to prod
using the config flag.
Once that is done, we'll remove `sync` mode to avoid the temporary code
duplication introduced by this PR.
The flag will be removed.
The `wait()` for the child process to exit is still synchronous; the
comment [here](
655d3b6468/pageserver/src/walredo.rs (L294-L306))
is still a valid argument in favor of that.
The `sync` mode had another implicit advantage: from tokio's
perspective, the calling task was using up coop budget.
But with `async` mode, that's no longer the case -- to tokio, the writes
to the child process pipe look like IO.
We could/should inform tokio about the CPU time budget consumed by the
task to achieve fairness similar to `sync`.
However, the [runtime function for this is
`tokio_unstable`](`https://docs.rs/tokio/latest/tokio/task/fn.consume_budget.html).
Refs
----
refs #6628
refs https://github.com/neondatabase/neon/issues/2975
No functional changes, this is a comments/naming PR.
While merging sharding changes, some cleanup of the shard.rs types was
deferred.
In this PR:
- Rename `is_zero` to `is_shard_zero` to make clear that this method
doesn't literally mean that the entire object is zeros, just that it
refers to the 0th shard in a tenant.
- Pull definitions of types to the top of shard.rs and add a big comment
giving an overview of which type is for what.
Closes: https://github.com/neondatabase/neon/issues/6072
This PR is a fallout from work on #7062.
# Changes
- Unify the freeze-and-flush and hard shutdown code paths into a single
method `Timeline::shutdown` that takes the shutdown mode as an argument.
- Replace `freeze_and_flush` bool arg in callers with that mode
argument, makes them more expressive.
- Switch timeline deletion to use `Timeline::shutdown` instead of its
own slightly-out-of-sync copy.
- Remove usage of `task_mgr::shutdown_watcher` /
`task_mgr::shutdown_token` where possible
# Future Work
Do we really need the freeze_and_flush?
If we could get rid of it, then there'd be no need for a specific
shutdown order.
Also, if you undo this patch's changes to the `eviction_task.rs` and
enable RUST_LOG=debug, it's easy to see that we do leave some task
hanging that logs under span `Connection{...}` at debug level. I think
it's a pre-existing issue; it's probably a broker client task.
## Problem
For reasons unrelated to this PR, I would like to make use of the tenant
conf in the `InMemoryLayer`. Previously, this was not possible without
copying and manually updating the copy to keep it in sync with updates.
## Summary of Changes:
Replace the `Arc<RwLock<AttachedTenantConf>>` with
`Arc<ArcSwap<AttachedTenantConf>>` (how many `Arc(s)` can one fit in a
type?). The most interesting part of this change is the updating of the
tenant config (`set_new_tenant_config` and
`set_new_location_config`). In theory, these two may race, although the
storage controller should prevent this via the tenant exclusive op lock.
Particular care has been taken to not "lose" a location config update by
using the read-copy-update approach when updating only the config.
We want to move the code base away from task_mgr.
This PR refactors the walreceiver code such that it doesn't use
`task_mgr` anymore.
# Background
As a reminder, there are three tasks in a Timeline that's ingesting WAL.
`WalReceiverManager`, `WalReceiverConnectionHandler`, and
`WalReceiverConnectionPoller`.
See the documentation in `task_mgr.rs` for how they interact.
Before this PR, cancellation was requested through
task_mgr::shutdown_token() and `TaskHandle::shutdown`.
Wait-for-task-finish was implemented using a mixture of
`task_mgr::shutdown_tasks` and `TaskHandle::shutdown`.
This drawing might help:
<img width="300" alt="image"
src="https://github.com/neondatabase/neon/assets/956573/b6be7ad6-ecb3-41d0-b410-ec85cb8d6d20">
# Changes
For cancellation, the entire WalReceiver task tree now has a
`child_token()` of `Timeline::cancel`. The `TaskHandle` no longer is a
cancellation root.
This means that `Timeline::cancel.cancel()` is propagated.
For wait-for-task-finish, all three tasks in the task tree hold the
`Timeline::gate` open until they exit.
The downside of using the `Timeline::gate` is that we can no longer wait
for just the walreceiver to shut down, which is particularly relevant
for `Timeline::flush_and_shutdown`.
Effectively, it means that we might ingest more WAL while the
`freeze_and_flush()` call is ongoing.
Also, drive-by-fix the assertiosn around task kinds in `wait_lsn`. The
check for `WalReceiverConnectionHandler` was ineffective because that
never was a task_mgr task, but a TaskHandle task. Refine the assertion
to check whether we would wait, and only fail in that case.
# Alternatives
I contemplated (ab-)using the `Gate` by having a separate `Gate` for
`struct WalReceiver`.
All the child tasks would use _that_ gate instead of `Timeline::gate`.
And `struct WalReceiver` itself would hold an `Option<GateGuard>` of the
`Timeline::gate`.
Then we could have a `WalReceiver::stop` function that closes the
WalReceiver's gate, then drops the `WalReceiver::Option<GateGuard>`.
However, such design would mean sharing the WalReceiver's `Gate` in an
`Arc`, which seems awkward.
A proper abstraction would be to make gates hierarchical, analogous to
CancellationToken.
In the end, @jcsp and I talked it over and we determined that it's not
worth the effort at this time.
# Refs
part of #7062
## Problem
Part of the legacy (but current) compaction algorithm is to find a stack
of overlapping delta layers which will be turned
into an image layer. This operation is exponential in terms of the
number of matching layers and we do it roughly every 20 seconds.
## Summary of changes
Only check if a new image layer is required if we've ingested a certain
amount of WAL since the last check.
The amount of wal is expressed in terms of multiples of checkpoint
distance, with the intuition being that
that there's little point doing the check if we only have two new L1
layers (not enough to create a new image).
## Problem
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
This change improves the resilience of the system to unclean restarts.
Previously, re-attach responses only included attached tenants
- If the pageserver had local state for a secondary location, it would
remain, but with no guarantee that it was still _meant_ to be there.
After this change, the pageserver will only retain secondary locations
if the /re-attach response indicates that they should still be there.
- If the pageserver had local state for an attached location that was
omitted from a re-attach response, it would be entirely detached. This
is wasteful in a typical HA setup, where an offline node's tenants might
have been re-attached elsewhere before it restarts, but the offline
node's location should revert to a secondary location rather than being
wiped. Including secondary tenants in the re-attach response enables the
pageserver to avoid throwing away local state unnecessarily.
In this PR:
- The re-attach items are extended with a 'mode' field.
- Storage controller populates 'mode'
- Pageserver interprets it (default is attached if missing) to construct
either a SecondaryTenant or a Tenant.
- A new test exercises both cases.
## Problem
If a shutdown happens when a tenant is attaching, we were logging at
ERROR severity and with a backtrace. Yuck.
## Summary of changes
- Pass a flag into `make_broken` to enable quietening this non-scary
case.
Switched the order; doing https://github.com/neondatabase/neon/pull/6139
first then can remove uninit marker after.
## Problem
Previously, existence of a timeline directory was treated as evidence of
the timeline's logical existence. That is no longer the case since we
treat remote storage as the source of truth on each startup: we can
therefore do without this mark file.
The mark file had also been used as a pseudo-lock to guard against
concurrent creations of the same TimelineId -- now that persistence is
no longer required, this is a bit unwieldy.
In #6139 the `Tenant::timelines_creating` was added to protect against
concurrent creations on the same TimelineId, making the uninit mark file
entirely redundant.
## Summary of changes
- Code that writes & reads mark file is removed
- Some nearby `pub` definitions are amended to `pub(crate)`
- `test_duplicate_creation` is added to demonstrate that mutual
exclusion of creations still works.
## Summary of changes
The problem it fixes is when `request_lsn` is `u64::MAX-1` the
`cont_lsn` becomes `u64::MAX` which is the same as `prev_lsn` which
stops the loop.
Closes https://github.com/neondatabase/neon/issues/6812
Tenant::shutdown or Timeline::shutdown completes and becomes externally
observable before the corresponding Tenant/Timeline object is dropped.
For example, after observing a Tenant::shutdown to complete, we could
attach the same tenant_id again. The shut down Tenant object might still
be around at the time of the attach.
The race is then the following:
- old object's metrics are still around
- new object uses with_label_values
- old object calls remove_label_values
The outcome is that the new object will have the metric objects (they're
an Arc internall) but the metrics won't be part of the internal registry
and hence they'll be missing in `/metrics`.
Later, when the new object gets shut down and tries to
remove_label_value, it will observe an error because
the metric was already removed by the old object.
Changes
-------
This PR moves metric removal to `shutdown()`.
An alternative design would be to multi-version the metrics using a
distinguishing label, or, to use a better metrics crate that allows
removing metrics from the registry through the locally held metric
handle instead of interacting with the (globally shared) registry.
refs https://github.com/neondatabase/neon/pull/7051
## Problem
We reverted https://github.com/neondatabase/neon/pull/6661 a few days
ago. The change led to OOMs in
benchmarks followed by large WAL reingests.
The issue was that we removed [this
code](d04af08567/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs (L409-L417)).
That call may trigger a roll of the open layer due to
the keepalive messages received from the safekeeper. Removing it meant
that enforcing
of checkpoint timeout became even more lax and led to using up large
amounts of memory
for the in memory layer indices.
## Summary of changes
Piggyback on keep alive messages to enforce checkpoint timeout. This is
a hack, but it's exactly what
the current code is doing.
## Alternatives
Christhian, Joonas and myself sketched out a timer based approach
[here](https://github.com/neondatabase/neon/pull/6940). While discussing
it further, it became obvious that's also a bit of a hack and not the
desired end state. I chose not
to take that further since it's not what we ultimately want and it'll be
harder to rip out.
Right now it's unclear what the ideal system behaviour is:
* early flushing on memory pressure, or ...
* detaching tenants on memory pressure
## Problem
When vectored get encountered a portion of the key range that could
not be mapped to any layer in the current timeline it would incorrectly
bail out of the current timeline. This is incorrect since we may have
had layers queued for a visit in the fringe.
## Summary of changes
* Add a repro unit test
* Remove the early bail out path
* Simplify range search return value
## Problem
- The storage controller is the source of truth for a tenant's stripe
size, but doesn't currently have a way to propagate that to compute:
we're just using the default stripe size everywhere.
Closes: https://github.com/neondatabase/neon/issues/6903
## Summary of changes
- Include stripe size in `ComputeHookNotifyRequest`
- Include stripe size in `LocationConfigResponse`
The stripe size is optional: it will only be advertised for
multi-sharded tenants. This enables the controller to defer the choice
of stripe size until we split a tenant for the first time.
Nightly has added a bunch of compiler and linter warnings. There is also
two dependencies that fail compilation on latest nightly due to using
the old `stdsimd` feature name. This PR fixes them.
Add off-by-default support for lazy queued tenant activation on attach.
This should be useful on bulk migrations as some tenants will be
activated faster due to operations or endpoint startup. Eventually all
tenants will get activated by reusing the same mechanism we have at
startup (`PageserverConf::concurrent_tenant_warmup`).
The difference to lazy attached tenants to startup ones is that we leave
their initial logical size calculation be triggered by WalReceiver or
consumption metrics.
Fixes: #6315
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
The vectored read path proposed in
https://github.com/neondatabase/neon/pull/6576 seems
to be functionally correct, but in my testing (see below) it is about 10-20% slower than the naive
sequential vectored implementation.
## Summary of changes
There's three parts to this PR:
1. Supporting vectored blob reads. This is actually trickier than it
sounds because on disk blobs are prefixed with a variable length size header.
Since the blobs are not necessarily fixed size, we need to juggle the offsets
such that the callers can retrieve the blobs from the resulting buffer.
2. Merge disk read requests issued by the vectored read path up to a
maximum size. Again, the merging is complicated by the fact that blobs
are not fixed size. We keep track of the begin and end offset of each blob
and pass them into the vectored blob reader. In turn, the reader will return
a buffer and the offsets at which the blobs begin and end.
3. A benchmark for basebackup requests against tenant with large SLRU
block counts is added. This required a small change to pagebench and a new config
variable for the pageserver which toggles the vectored get validation.
We can probably optimise things further by adding a little bit of
concurrency for our IO. In principle, it's as simple as spawning a task which deals with issuing
IO and doing the serialisation and handling on the parent task which receives input via a
channel.
This reverts commits 587cb705b8 (PR #6661)
and fcbe9fb184 (PR #6842).
Conflicts:
pageserver/src/tenant.rs
pageserver/src/tenant/timeline.rs
The conflicts were with
* pageserver: adjust checkpoint distance for sharded tenants (#6852)
* pageserver: add vectored get implementation (#6576)
Also we had to keep the `allowed_errors` to make `test_forward_compatibility` happy,
see the PR thread on GitHub for details.
Rebased version of #5234, part of #6768
This consists of three parts:
1. A refactoring and new contract for implementing and testing
compaction.
The logic is now in a separate crate, with no dependency on the
'pageserver' crate. It defines an interface that the real pageserver
must implement, in order to call the compaction algorithm. The interface
models things like delta and image layers, but just the parts that the
compaction algorithm needs to make decisions. That makes it easier unit
test the algorithm and experiment with different implementations.
I did not convert the current code to the new abstraction, however. When
compaction algorithm is set to "Legacy", we just use the old code. It
might be worthwhile to convert the old code to the new abstraction, so
that we can compare the behavior of the new algorithm against the old
one, using the same simulated cases. If we do that, have to be careful
that the converted code really is equivalent to the old.
This inclues only trivial changes to the main pageserver code. All the
new code is behind a tenant config option. So this should be pretty safe
to merge, even if the new implementation is buggy, as long as we don't
enable it.
2. A new compaction algorithm, implemented using the new abstraction.
The new algorithm is tiered compaction. It is inspired by the PoC at PR
#4539, although I did not use that code directly, as I needed the new
implementation to fit the new abstraction. The algorithm here is less
advanced, I did not implement partial image layers, for example. I
wanted to keep it simple on purpose, so that as we add bells and
whistles, we can see the effects using the included simulator.
One difference to #4539 and your typical LSM tree implementations is how
we keep track of the LSM tree levels. This PR doesn't have a permanent
concept of a level, tier or sorted run at all. There are just delta and
image layers. However, when compaction starts, we look at the layers
that exist, and arrange them into levels, depending on their shapes.
That is ephemeral: when the compaction finishes, we forget that
information. This allows the new algorithm to work without any extra
bookkeeping. That makes it easier to transition from the old algorithm
to new, and back again.
There is just a new tenant config option to choose the compaction
algorithm. The default is "Legacy", meaning the current algorithm in
'main'. If you set it to "Tiered", the new algorithm is used.
3. A simulator, which implements the new abstraction.
The simulator can be used to analyze write and storage amplification,
without running a test with the full pageserver. It can also draw an SVG
animation of the simulation, to visualize how layers are created and
deleted.
To run the simulator:
cargo run --bin compaction-simulator run-suite
---------
Co-authored-by: Heikki Linnakangas <heikki@neon.tech>
Noticed that we are failing to handle `Result::Err` when entering a gate
for logical size calculation. Audited rest of the gate enters, which
seem fine, unified two instances.
Noticed that the gate guard allows to remove a failpoint, then noticed
that adjacent failpoint was blocking the executor thread instead of
using `pausable_failpoint!`, fix both.
eviction_task.rs now maintains a gate guard as well.
Cc: #4733
## Problem
Previously we always wrote out both legacy and modern tenant config
files. The legacy write enabled rollbacks, but we are long past the
point where that is needed.
We still need the legacy format for situations where someone is running
tenants without generations (that will be yanked as well eventually),
but we can avoid writing it out at all if we do have a generation number
set. We implicitly also avoid writing the legacy config if our mode is
Secondary (secondary mode is newer than generations).
## Summary of changes
- Make writing legacy tenant config conditional on there being no
generation number set.
It's been dead-code-at-runtime for 9 months, let's remove it.
We can always re-introduce it at a later point.
Came across this while working on #6861, which will touch
`time_for_new_image_layer`. This is an opporunity to make that function
simpler.
Reverts neondatabase/neon#6765 , bringing back #6731
We concluded that #6731 never was the root cause for the instability in
staging.
More details:
https://neondb.slack.com/archives/C033RQ5SPDH/p1708011674755319
However, the massive amount of concurrent `spawn_blocking` calls from
the `save_metadata` calls during startups might cause a performance
regression.
So, we'll merge this PR here after we've stopped writing the metadata
#6769).
This PR introduces a new vectored implementation of the read path.
The search is basically a DFS if you squint at it long enough.
LayerFringe tracks the next layers to visit and acts as our stack.
Vertices are tuples of (layer, keyspace, lsn range). Continuously
pop the top of the stack (most recent layer) and do all the reads
for one layer at once.
The search maintains a fringe (`LayerFringe`) which tracks all the
layers that intersect the current keyspace being searched. Continuously
pop the top of the fringe (layer with highest LSN) and get all the data
required from the layer in one go.
Said search is done on one timeline at a time. If data is still required for
some keys, then search the ancestor timeline.
Apart from the high level layer traversal, vectored variants have been
introduced for grabbing data from each layer type. They still suffer from
read amplification issues and that will be addressed in a different PR.
You might notice that in some places we duplicate the code for the
existing read path. All of that code will be removed when we switch
the non-vectored read path to proxy into the vectored read path.
In the meantime, we'll have to contend with the extra cruft for the sake
of testing and gentle releasing.
## Problem
One WAL record can actually produce an arbitrary amount of key value pairs.
This is problematic since it might cause our frozen layers to bloat past the
max allowed size of S3 single shot uploads.
[#6639](https://github.com/neondatabase/neon/pull/6639) introduced a "should roll"
check after every batch of `ingest_batch_size` (100 WAL records by default). This helps,
but the original problem still exists.
## Summary of changes
This patch moves the responsibility of rolling the currently open layer
to the `TimelineWriter`. Previously, this was done ad-hoc via calls
to `check_checkpoint_distance`. The advantages of this approach are:
* ability to split one batch over multiple open layers
* less layer map locking
* remove ad-hoc check_checkpoint_distance calls
More specifically, we track the current size of the open layer in the
writer. On each `put` check whether the current layer should be closed
and a new one opened. Keeping track of the currently open layer results
in less contention on the layer map lock. It only needs to be acquired
on the first write and on writes that require a roll afterwards.
Rolling the open layer can be triggered by:
1. The distance from the last LSN we rolled at. This bounds the amount
of WAL that the safekeepers need to store.
2. The size of the currently open layer.
3. The time since the last roll. It helps safekeepers to regard
pageserver as caught up and suspend activity.
Closes#6624
## Problem
The ShardCount type has a magic '0' value that represents a legacy
single-sharded tenant, whose TenantShardId is formatted without a
`-0001` suffix (i.e. formatted as a traditional TenantId).
This was error-prone in code locations that wanted the actual number of
shards: they had to handle the 0 case specially.
## Summary of changes
- Make the internal value of ShardCount private, and expose `count()`
and `literal()` getters so that callers have to explicitly say whether
they want the literal value (e.g. for storing in a TenantShardId), or
the actual number of shards in the tenant.
---------
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
Cancellation and timeouts are handled at remote_storage callsites, if
they are. However they should always be handled, because we've had
transient problems with remote storage connections.
- Add cancellation token to the `trait RemoteStorage` methods
- For `download*`, `list*` methods there is
`DownloadError::{Cancelled,Timeout}`
- For the rest now using `anyhow::Error`, it will have root cause
`remote_storage::TimeoutOrCancel::{Cancel,Timeout}`
- Both types have `::is_permanent` equivalent which should be passed to
`backoff::retry`
- New generic RemoteStorageConfig option `timeout`, defaults to 120s
- Start counting timeouts only after acquiring concurrency limiter
permit
- Cancellable permit acquiring
- Download stream timeout or cancellation is communicated via an
`std::io::Error`
- Exit backoff::retry by marking cancellation errors permanent
Fixes: #6096Closes: #4781
Co-authored-by: arpad-m <arpad-m@users.noreply.github.com>
## Problem
Aux files were stored with an O(N^2) cost, since on each modification
the entire map is re-written as a page image.
This addresses one axis of the inefficiency in logical replication's use
of storage (https://github.com/neondatabase/neon/issues/6626). It will
still be writing a large amount of duplicative data if writing the same
slot's state every 15 seconds, but the impact will be O(N) instead of
O(N^2).
## Summary of changes
- Introduce `NeonWalRecord::AuxFile`
- In `DatadirModification`, if the AUX_FILES_KEY has already been set,
then write a delta instead of an image
Some callers of `VirtualFile::crashsafe_overwrite` call it on the
executor thread, thereby potentially stalling it.
Others are more diligent and wrap it in `spawn_blocking(...,
Handle::block_on, ... )` to avoid stalling the executor thread.
However, because `crashsafe_overwrite` uses
VirtualFile::open_with_options internally, we spawn a new thread-local
`tokio-epoll-uring::System` in the blocking pool thread that's used for
the `spawn_blocking` call.
This PR refactors the situation such that we do the `spawn_blocking`
inside `VirtualFile::crashsafe_overwrite`. This unifies the situation
for the better:
1. Callers who didn't wrap in `spawn_blocking(..., Handle::block_on,
...)` before no longer stall the executor.
2. Callers who did it before now can avoid the `block_on`, resolving the
problem with the short-lived `tokio-epoll-uring::System`s in the
blocking pool threads.
A future PR will build on top of this and divert to tokio-epoll-uring if
it's configures as the IO engine.
Changes
-------
- Convert implementation to std::fs and move it into `crashsafe.rs`
- Yes, I know, Safekeepers (cc @arssher ) added `durable_rename` and
`fsync_async_opt` recently. However, `crashsafe_overwrite` is different
in the sense that it's higher level, i.e., it's more like
`std::fs::write` and the Safekeeper team's code is more building block
style.
- The consequence is that we don't use the VirtualFile file descriptor
cache anymore.
- I don't think it's a big deal because we have plenty of slack wrt
production file descriptor limit rlimit (see [this
dashboard](https://neonprod.grafana.net/d/e4a40325-9acf-4aa0-8fd9-f6322b3f30bd/pageserver-open-file-descriptors?orgId=1))
- Use `tokio::task::spawn_blocking` in
`VirtualFile::crashsafe_overwrite` to call the new
`crashsafe::overwrite` API.
- Inspect all callers to remove any double-`spawn_blocking`
- spawn_blocking requires the captures data to be 'static + Send. So,
refactor the callers. We'll need this for future tokio-epoll-uring
support anyway, because tokio-epoll-uring requires owned buffers.
Related Issues
--------------
- overall epic to enable write path to tokio-epoll-uring: #6663
- this is also kind of relevant to the tokio-epoll-uring System creation
failures that we encountered in staging, investigation being tracked in
#6667
- why is it relevant? Because this PR removes two uses of
`spawn_blocking+Handle::block_on`
The smaller changes I found while looking around #6584.
- rustfmt was not able to format handle_timeline_create
- fix Generation::get_suffix always allocating
- Generation was missing a `#[track_caller]` for panicky method
- attach has a lot of issues, but even with this PR it cannot be
formatted by rustfmt
- moved the `preload` span to be on top of `attach` -- it is awaited
inline
- make disconnected panic! or unreachable! into expect, expect_err
This PR is preliminary cleanups and refactoring around `remote_storage`
for next PR which will move the timeouts and cancellation into
`remote_storage`.
Summary:
- smaller drive-by fixes
- code simplification
- refactor common parts like `DownloadError::is_permanent`
- align error types with `RemoteStorage::list_*` to use more
`download_retry` helper
Cc: #6096
## Problem
One doesn't know at tenant creation time how large the tenant will grow.
We need to be able to dynamically adjust the shard count at runtime.
This is implemented as "splitting" of shards into smaller child shards,
which cover a subset of the keyspace that the parent covered.
Refer to RFC: https://github.com/neondatabase/neon/pull/6358
Part of epic: #6278
## Summary of changes
This PR implements the happy path (does not cleanly recover from a crash
mid-split, although won't lose any data), without any optimizations
(e.g. child shards re-download their own copies of layers that the
parent shard already had on local disk)
- Add `/v1/tenant/:tenant_shard_id/shard_split` API to pageserver: this
copies the shard's index to the child shards' paths, instantiates child
`Tenant` object, and tears down parent `Tenant` object.
- Add `splitting` column to `tenant_shards` table. This is written into
an existing migration because we haven't deployed yet, so don't need to
cleanly upgrade.
- Add `/control/v1/tenant/:tenant_id/shard_split` API to
attachment_service,
- Add `test_sharding_split_smoke` test. This covers the happy path:
future PRs will add tests that exercise failure cases.
When we'll later introduce a global pool of pre-spawned walredo
processes (https://github.com/neondatabase/neon/issues/6581), this
refactoring avoids plumbing through the reference to the pool to all the
places where we create a broken tenant.
Builds atop the refactoring in #6583
The solution we ended up for `backoff::retry` requires always cloning of
cancellation tokens even though there is just `.await`. Fix that, and
also turn the return type into `Option<Result<T, E>>` avoiding the need
for the `E::cancelled()` fn passed in.
Cc: #6096
Before tenant migration it made sense to leak broken tenants in the
metrics until restart. Nowdays it makes less sense because on
cancellations we set the tenant broken. The set metric still allows
filterable alerting.
Fixes: #6507
