# Refs
- fixes https://github.com/neondatabase/neon/issues/10309
- fixup of batching design, first introduced in
https://github.com/neondatabase/neon/pull/9851
- refinement of https://github.com/neondatabase/neon/pull/8339
# Problem
`Tenant::shutdown` was occasionally taking many minutes (sometimes up to
20) in staging and prod if the
`page_service_pipelining.mode="concurrent-futures"` is enabled.
# Symptoms
The issue happens during shard migration between pageservers.
There is page_service unavailability and hence effectively downtime for
customers in the following case:
1. The source (state `AttachedStale`) gets stuck in `Tenant::shutdown`,
waiting for the gate to close.
2. Cplane/Storcon decides to transition the target `AttachedMulti` to
`AttachedSingle`.
3. That transition comes with a bump of the generation number, causing
the `PUT .../location_config` endpoint to do a full `Tenant::shutdown` /
`Tenant::attach` cycle for the target location.
4. That `Tenant::shutdown` on the target gets stuck, waiting for the
gate to close.
5. Eventually the gate closes (`close completed`), correlating with a
`page_service` connection handler logging that it's exiting because of a
network error (`Connection reset by peer` or `Broken pipe`).
While in (4):
- `Tenant::shutdown` is stuck waiting for all `Timeline::shutdown` calls
to complete.
So, really, this is a `Timeline::shutdown` bug.
- retries from Cplane/Storcon to complete above state transitions, fail
with errors related to the tenant mgr slot being in state
`TenantSlot::InProgress`, the tenant state being
`TenantState::Stopping`, and the timelines being in
`TimelineState::Stopping`, and the `Timeline::cancel` being cancelled.
- Existing (and/or new?) page_service connections log errors `error
reading relation or page version: Not found: Timed out waiting 30s for
tenant active state. Latest state: None`
# Root-Cause
After a lengthy investigation ([internal
write-up](https://www.notion.so/neondatabase/2025-01-09-batching-deadlock-Slow-Log-Analysis-in-Staging-176f189e00478050bc21c1a072157ca4?pvs=4))
I arrived at the following root cause.
The `spsc_fold` channel (`batch_tx`/`batch_rx`) that connects the
Batcher and Executor stages of the pipelined mode was storing a `Handle`
and thus `GateGuard` of the Timeline that was not shutting down.
The design assumption with pipelining was that this would always be a
short transient state.
However, that was incorrect: the Executor was stuck on writing/flushing
an earlier response into the connection to the client, i.e., socket
write being slow because of TCP backpressure.
The probable scenario of how we end up in that case:
1. Compute backend process sends a continuous stream of getpage prefetch
requests into the connection, but never reads the responses (why this
happens: see Appendix section).
2. Batch N is processed by Batcher and Executor, up to the point where
Executor starts flushing the response.
3. Batch N+1 is procssed by Batcher and queued in the `spsc_fold`.
4. Executor is still waiting for batch N flush to finish.
5. Batcher eventually hits the `TimeoutReader` error (10min).
From here on it waits on the
`spsc_fold.send(Err(QueryError(TimeoutReader_error)))`
which will never finish because the batch already inside the `spsc_fold`
is not
being read by the Executor, because the Executor is still stuck in the
flush.
(This state is not observable at our default `info` log level)
6. Eventually, Compute backend process is killed (`close()` on the
socket) or Compute as a whole gets killed (probably no clean TCP
shutdown happening in that case).
7. Eventually, Pageserver TCP stack learns about (6) through RST packets
and the Executor's flush() call fails with an error.
8. The Executor exits, dropping `cancel_batcher` and its end of the
spsc_fold.
This wakes Batcher, causing the `spsc_fold.send` to fail.
Batcher exits.
The pipeline shuts down as intended.
We return from `process_query` and log the `Connection reset by peer` or
`Broken pipe` error.
The following diagram visualizes the wait-for graph at (5)
```mermaid
flowchart TD
Batcher --spsc_fold.send(TimeoutReader_error)--> Executor
Executor --flush batch N responses--> socket.write_end
socket.write_end --wait for TCP window to move forward--> Compute
```
# Analysis
By holding the GateGuard inside the `spsc_fold` open, the pipelining
implementation
violated the principle established in
(https://github.com/neondatabase/neon/pull/8339).
That is, that `Handle`s must only be held across an await point if that
await point
is sensitive to the `<Handle as Deref<Target=Timeline>>::cancel` token.
In this case, we were holding the Handle inside the `spsc_fold` while
awaiting the
`pgb_writer.flush()` future.
One may jump to the conclusion that we should simply peek into the
spsc_fold to get
that Timeline cancel token and be sensitive to it during flush, then.
But that violates another principle of the design from
https://github.com/neondatabase/neon/pull/8339.
That is, that the page_service connection lifecycle and the Timeline
lifecycles must be completely decoupled.
Tt must be possible to shut down one shard without shutting down the
page_service connection, because on that single connection we might be
serving other shards attached to this pageserver.
(The current compute client opens separate connections per shard, but,
there are plans to change that.)
# Solution
This PR adds a `handle::WeakHandle` struct that does _not_ hold the
timeline gate open.
It must be `upgrade()`d to get a `handle::Handle`.
That `handle::Handle` _does_ hold the timeline gate open.
The batch queued inside the `spsc_fold` only holds a `WeakHandle`.
We only upgrade it while calling into the various `handle_` methods,
i.e., while interacting with the `Timeline` via `<Handle as
Deref<Target=Timeline>>`.
All that code has always been required to be (and is!) sensitive to
`Timeline::cancel`, and therefore we're guaranteed to bail from it
quickly when `Timeline::shutdown` starts.
We will drop the `Handle` immediately, before we start
`pgb_writer.flush()`ing the responses.
Thereby letting go of our hold on the `GateGuard`, allowing the timeline
shutdown to complete while the page_service handler remains intact.
# Code Changes
* Reproducer & Regression Test
* Developed and proven to reproduce the issue in
https://github.com/neondatabase/neon/pull/10399
* Add a `Test` message to the pagestream protocol (`cfg(feature =
"testing")`).
* Drive-by minimal improvement to the parsing code, we now have a
`PagestreamFeMessageTag`.
* Refactor `pageserver/client` to allow sending and receiving
`page_service` requests independently.
* Add a Rust helper binary to produce situation (4) from above
* Rationale: (4) and (5) are the same bug class, we're holding a gate
open while `flush()`ing.
* Add a Python regression test that uses the helper binary to
demonstrate the problem.
* Fix
* Introduce and use `WeakHandle` as explained earlier.
* Replace the `shut_down` atomic with two enum states for `HandleInner`,
wrapped in a `Mutex`.
* To make `WeakHandle::upgrade()` and `Handle::downgrade()`
cache-efficient:
* Wrap the `Types::Timeline` in an `Arc`
* Wrap the `GateGuard` in an `Arc`
* The separate `Arc`s enable uncontended cloning of the timeline
reference in `upgrade()` and `downgrade()`.
If instead we were `Arc<Timeline>::clone`, different connection handlers
would be hitting the same cache line on every upgrade()/downgrade(),
causing contention.
* Please read the udpated module-level comment in `mod handle`
module-level comment for details.
# Testing & Performance
The reproducer test that failed before the changes now passes, and
obviously other tests are passing as well.
We'll do more testing in staging, where the issue happens every ~4h if
chaos migrations are enabled in storcon.
Existing perf testing will be sufficient, no perf degradation is
expected.
It's a few more alloctations due to the added Arc's, but, they're low
frequency.
# Appendix: Why Compute Sometimes Doesn't Read Responses
Remember, the whole problem surfaced because flush() was slow because
Compute was not reading responses. Why is that?
In short, the way the compute works, it only advances the page_service
protocol processing when it has an interest in data, i.e., when the
pagestore smgr is called to return pages.
Thus, if compute issues a bunch of requests as part of prefetch but then
it turns out it can service the query without reading those pages, it
may very well happen that these messages stay in the TCP until the next
smgr read happens, either in that session, or possibly in another
session.
If there’s too many unread responses in the TCP, the pageserver kernel
is going to backpressure into userspace, resulting in our stuck flush().
All of this stems from the way vanilla Postgres does prefetching and
"async IO":
it issues `fadvise()` to make the kernel do the IO in the background,
buffering results in the kernel page cache.
It then consumes the results through synchronous `read()` system calls,
which hopefully will be fast because of the `fadvise()`.
If it turns out that some / all of the prefetch results are not needed,
Postgres will not be issuing those `read()` system calls.
The kernel will eventually react to that by reusing page cache pages
that hold completed prefetched data.
Uncompleted prefetch requests may or may not be processed -- it's up to
the kernel.
In Neon, the smgr + Pageserver together take on the role of the kernel
in above paragraphs.
In the current implementation, all prefetches are sent as GetPage
requests to Pageserver.
The responses are only processed in the places where vanilla Postgres
would do the synchronous `read()` system call.
If we never get to that, the responses are queued inside the TCP
connection, which, once buffers run full, will backpressure into
Pageserver's sending code, i.e., the `pgb_writer.flush()` that was the
root cause of the problems we're fixing in this PR.
## Problem
All pageserver have the same application name which makes it hard to
distinguish them.
## Summary of changes
Include the node id in the application name sent to the safekeeper. This
should gives us
more visibility in logs. There's a few metrics that will increase in
cardinality by `pageserver_count`,
but that's fine.
## Problem
gc-compaction needs the partitioning data to decide the job split. This
refactor allows concurrent access/computing the partitioning.
## Summary of changes
Make `partitioning` an ArcSwap so that others can access the
partitioning while we compute it. Fully eliminate the `repartition is
called concurrently` warning when gc-compaction is going on.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
part of https://github.com/neondatabase/neon/issues/9114
part of investigation of
https://github.com/neondatabase/neon/issues/10049
## Summary of changes
* If `cfg!(test) or cfg!(feature = testing)`, then we will always try
generating an image to ensure the history is replayable, but not put the
image layer into the final layer results, therefore discovering wrong
key history before we hit a read error.
* I suspect it's easier to trigger some races if gc-compaction is
continuously run on a timeline, so I increased the frequency to twice
per 10 churns.
* Also, create branches in gc-compaction smoke tests to get more test
coverage.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Arpad Müller <arpad@neon.tech>
## Problem
When a pageserver is receiving high rates of requests, we don't have a
good way to efficiently discover what the client's access pattern is.
Closes: https://github.com/neondatabase/neon/issues/10275
## Summary of changes
- Add
`/v1/tenant/x/timeline/y/page_trace?size_limit_bytes=...&time_limit_secs=...`
API, which returns a binary buffer.
- Add `pagectl page-trace` tool to decode and analyze the output.
---------
Co-authored-by: Erik Grinaker <erik@neon.tech>
## Problem
Currently, we call `InterpretedWalRecord::from_bytes_filtered`
from each shard. To serve multiple shards at the same time,
the API needs to allow for enquiring about multiple shards.
## Summary of changes
This commit tweaks it a pretty brute force way. Naively, we could
just generate the shard for a key, but pre and post split shards
may be subscribed at the same time, so doing it efficiently is more
complex.
## 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.
## Problem
Before this PR, the pagestream throttle was applied weighted on a
per-batch basis.
This had several problems:
1. The throttle occurence counters were only bumped by `1` instead of
`batch_size`.
2. The throttle wait time aggregator metric only counted one wait time,
irrespective
of `batch_size`. That makes sense in some ways of looking at it but not
in others.
3. If the last request in the batch runs into the throttle, the other
requests in the
batch are also throttled, i.e., over-throttling happens (theoretical,
didn't measure
it in practice).
## Solution
It occured to me that we can simply push the throttling upwards into
`pagestream_read_message`.
This has the added benefit that in pipeline mode, the `executor` stage
will, if it is idle,
steal whatever requests already made it into the `spsc_fold` and execute
them; before this
change, that was not the case - the throttling happened in the
`executor` stage instead of
the `batcher` stage.
## Code Changes
There are two changes in this PR:
1. Lifting up the throttling into the `pagestream_read_message` method.
2. Move the throttling metrics out of the `Throttle` type into
`SmgrOpMetrics`.
Unlike the other smgr metrics, throttling is per-tenant, hence the Arc.
3. Refactor the `SmgrOpTimer` implementation to account for the new
observation states,
and simplify its design.
4. Drive-by-fix flush time metrics. It was using the same `now` in the
`observe_guard` every time.
The `SmgrOpTimer` is now a state machine.
Each observation point moves the state machine forward.
If a timer object is dropped early some "pair"-like metrics still
require an increment or observation.
That's done in the Drop implementation, by driving the state machine to
completion.
## Problem
Discovered during the relation dir refactor work.
If we do not create images as in this patch, we would get two set of
image layers:
```
0000...METADATA_KEYS
0000...REL_KEYS
```
They overlap at the same LSN and would cause data loss for relation
keys. This doesn't happen in prod because initial image layer generation
is never called, but better to be fixed to avoid future issues with the
reldir refactors.
## Summary of changes
* Consolidate create_image_layers call into a single one.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
In preparation to https://github.com/neondatabase/neon/issues/9516. We
need to store rel size and directory data in the sparse keyspace, but it
does not support inheritance yet.
## Summary of changes
Add a new type of keyspace "sparse but inherited" into the system.
On the read path: we don't remove the key range when we descend into the
ancestor. The search will stop when (1) the full key range is covered by
image layers (which has already been implemented before), or (2) we
reach the end of the ancestor chain.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
With a new beta build of the rust compiler, it's good to check out the
new lints. Either to find false positives, or find flaws in our code.
Additionally, it helps reduce the effort required to update to 1.85 in 6
weeks.
## Problem
Currently, the heap profiling frequency is every 1 MB allocated. Taking
a profile stack trace takes about 1 µs, and allocating 1 MB takes about
15 µs, so the overhead is about 6.7% which is a bit high. This is a
fixed cost regardless of whether heap profiles are actually accessed.
## Summary of changes
Increase the heap profiling sample frequency from 1 MB to 2 MB, which
reduces the overhead to about 3.3%. This seems acceptable, considering
performance-sensitive code will avoid allocations as far as possible
anyway.
When we moved throttling up from Timeline::get into page_service,
we stopped being sensitive to `Timeline::cancel`, even though we're
holding a Handle and thus a guard on the `Timeline::gate` open.
This PR rectifies the situation.
Refs
- Found while investigating #10309 (hung detach because gate kept open),
but not expected to be the root cause of that issue because the
affected tenants are not being throttled according to their metrics.
## Problem
Auto-offloading as requested by the compaction task is racy with
unarchival, in that the compaction task might attempt to offload an
unarchived timeline. By that point it will already have set the timeline
to the `Stopping` state however, which makes it unusable for any
purpose. For example:
1. compaction task decides to offload timeline
2. timeline gets unarchived
3. `offload_timeline` gets called by compaction task
* sets timeline's state to `Stopping`
* realizes that the timeline can't be unarchived, errors out
6. endpoint can't be started as the timeline is `Stopping` and thus
'can't be found'.
A future iteration of the compaction task can't "heal" this state either
as the timeline will still not be archived, same goes for other
automatic stuff. The only way to heal this is a tenant detach+attach, or
alternatively a pageserver restart.
Furthermore, the compaction task is especially amenable for such races
as it first stores `can_offload` into a variable, figures out whether
compaction is needed (which takes some time), and only then does it
attempt an offload operation: the time difference between "check" and
"use" is non-trivially small.
To make it even worse, we start the compaction task right after attach
of a tenant, and it is a common pattern by pageserver users to attach a
tenant to then immediately unarchive a timeline, so that an endpoint can
be started.
## Solutions not adopted
The simplest solution is to move the `can_offload` check to right before
attempting of the offload. But this is not a good solution, as no lock
is held between that check and timeline shutdown. So races would still
be possible, just become less likely.
I explored using the timeline state for this, as in adding an additional
enum variant. But `Timeline::set_state` is racy (#10297).
## Adopted solution
We use the lock on the timeline's upload queue as an arbiter: either
unarchival gets to it first and sours the state for auto-offloading, or
auto-offloading shuts it down, which stops any parallel unarchival in
its tracks. The key part is not releasing the upload queue's lock
between the check whether the timeline is archived or not, and shutting
it down (the actual implementation only sets `shutting_down` but it has
the same effect on `initialized_mut()` as a full shutdown). The rest of
the patch is stuff that follows from this.
We also move the part where we set the state to `Stopping` to after that
arbiter has decided the fate of the timeline. For deletions, we do keep
it inside `DeleteTimelineFlow::prepare` however, so that it is called
with all of the the timelines locks held that the function allocates
(timelines lock most importantly). This is only a precautionary measure
however, as I didn't want to analyze deletion related code for possible
races.
## Future changes
It might make sense to move `can_offload` to right before the offload
attempt. Maybe some other properties might have changed as well.
Although this will not be perfect either as no lock is held. I want to
keep it out of this change to emphasize that this move wasn't the main
reason we are race free now.
Fixes#10220
## Problem
If for some reasons we already garbage-collected the data under an LSN
but the caller uses a past LSN for the find_time_cutoff function, now we
will report a missing key error and GC will never proceed.
Note that missing key error can also happen if the key is really missing
(i.e., during the past offload incidents)
## Summary of changes
Make sure GC proceeds by bumping the LSN. When time_cutoff=None, we will
not increase the time_cutoff (it will be set to latest_gc_cutoff). If we
really need to bump the GC LSN for maintenance purpose, we need a
separate API to do that.
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
We have several serious data corruption incidents caused by mismatch of
get-age requests:
https://neondb.slack.com/archives/C07FJS4QF7V/p1723032720164359
We hope that the problem is fixed now. But it is better to prevent such
kind of problems in future.
Part of https://github.com/neondatabase/cloud/issues/16472
## Summary of changes
This PR introduce new V3 version of compute<->pageserver protocol,
adding tag to getpage response.
So now compute is able to check if it really gets response to the
requested page.
## 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
---------
Co-authored-by: Konstantin Knizhnik <knizhnik@neon.tech>
Co-authored-by: Heikki Linnakangas <heikki@neon.tech>
## Problem
The filtered record metric doesn't make sense for interpreted ingest.
## Summary of changes
While of dubious utility in the first place, this patch replaces them
with records received and records observed metrics for interpreted
ingest:
* received records cause the pageserver to do _something_: write a key,
value pair to storage, update some metadata or flush pending
modifications
* observed records are a shard 0 concept and contain only key metadata
used in tracking relation sizes (received records include observed
records)
## Problem
close https://github.com/neondatabase/neon/issues/10192
## Summary of changes
* `find_gc_time_cutoff` takes `now` parameter so that all branches
compute the cutoff based on the same start time, avoiding races.
* gc-compaction uses a single `get_gc_compaction_watermark` function to
get the safe LSN to compact.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
It's not legal to modify layers that are referenced by the current layer
index. Assert this in the upload queue, as preparation for upload queue
reordering.
Touches #10096.
## Summary of changes
Add a debug assertion that the upload queue does not modify layers
referenced by the current index.
I could be convinced that this should be a plain assertion, but will be
conservative for now.
## Problem
Since enabling continuous profiling in staging, we've seen frequent seg
faults. This is suspected to be because jemalloc and pprof-rs take a
stack trace at the same time, and the handlers aren't signal safe.
jemalloc does this probabilistically on every allocation, regardless of
whether someone is taking a heap profile, which means that any CPU
profile has a chance to cause a seg fault.
Touches #10225.
## Summary of changes
For now, just disable heap profiles -- CPU profiles are more important,
and we need to be able to take them without risking a crash.
There is a race condition between `Tenant::shutdown`'s `defuse_for_drop`
loop and `offload_timeline`, where timeline offloading can insert into a
tenant that is in the process of shutting down, in fact so far
progressed that the `defuse_for_drop` has already been called.
This prevents warn log lines of the form:
```
offloaded timeline <hash> was dropped without having cleaned it up at the ancestor
```
The solution piggybacks on the `offloaded_timelines` lock: both the
defuse loop and the offloaded timeline insertion need to acquire the
lock, and we know that the defuse loop only runs after the tenant has
set its `TenantState` to `Stopping`.
So if we hold the `offloaded_timelines` lock, and know that the
`TenantState` is not `Stopping`, then we know that the defuse loop has
not ran yet, and holding the lock ensures that it doesn't start running
while we are inserting the offloaded timeline.
Fixes#10070
## Problem
In https://github.com/neondatabase/neon/pull/9897 we temporarily
disabled the layer valid check because the current one only considers
the end result of all compaction algorithms, but partial gc-compaction
would temporarily produce an "invalid" layer map.
part of https://github.com/neondatabase/neon/issues/9114
## Summary of changes
Allow LSN splits to overlap in the slow path check. Currently, the valid
check is only used in storage scrubber (background job) and during
gc-compaction (without taking layer lock). Therefore, it's fine for such
checks to be a little bit inefficient but more accurate.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
We cannot get the size of the compaction queue and access the info.
Part of #9114
## Summary of changes
* Add an API endpoint to get the compaction queue.
* gc_compaction test case now waits until the compaction finishes.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
part of https://github.com/neondatabase/neon/issues/9114
In https://github.com/neondatabase/neon/pull/10127 we fixed the race,
but we didn't add the errors to the allowlist.
## Summary of changes
* Allow repartition errors in the gc-compaction smoke test.
I think it might be worth to refactor the code to allow multiple threads
getting a copy of repartition status (i.e., using Rcu) in the future.
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
Changes in #9786 were functionally complete but missed some edges that
made testing less robust than it should have been:
- `is_key_disposable` didn't consider SLRU dir keys disposable
- Timeline `init_empty` was always creating SLRU dir keys on all shards
The result was that when we had a bug
(https://github.com/neondatabase/neon/pull/10080), it wasn't apparent in
tests, because one would only encounter the issue if running on a
long-lived timeline with enough compaction to drop the initially created
empty SLRU dir keys, _and_ some CLog truncation going on.
Closes: https://github.com/neondatabase/cloud/issues/21516
## Summary of changes
- Update is_key_global and init_empty to handle SLRU dir keys properly
-- the only functional impact is that we avoid writing some spurious
keys in shards >0, but this makes testing much more robust.
- Make `test_clog_truncate` explicitly use a sharded tenant
The net result is that if one reverts #10080, then tests fail (i.e. this
PR is a reproducer for the issue)
## Problem
In #8550, we made the flush loop wait for uploads after every layer.
This was to avoid unbounded buildup of uploads, and to reduce compaction
debt. However, the approach has several problems:
* It prevents upload parallelism.
* It prevents flush and upload pipelining.
* It slows down ingestion even when there is no need to backpressure.
* It does not directly backpressure WAL ingestion (only via
`disk_consistent_lsn`), and will build up in-memory layers.
* It does not directly backpressure based on compaction debt and read
amplification.
An alternative solution to these problems is proposed in #8390.
In the meanwhile, we revert the change to reduce the impact on ingest
throughput. This does reintroduce some risk of unbounded
upload/compaction buildup. Until
https://github.com/neondatabase/neon/issues/8390, this can be addressed
in other ways:
* Use `max_replication_apply_lag` (aka `remote_consistent_lsn`), which
will more directly limit upload debt.
* Shard the tenant, which will spread the flush/upload work across more
Pageservers and move the bottleneck to Safekeeper.
Touches #10095.
## Summary of changes
Remove waiting on the upload queue in the flush loop.
## Problem
close https://github.com/neondatabase/neon/issues/10124
gc-compaction split_gc_jobs is holding the repartition lock for too long
time.
## Summary of changes
* Ensure split_gc_compaction_jobs drops the repartition lock once it
finishes cloning the structures.
* Update comments.
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
Cplane and storage controller tenant config changes are not additive.
Any change overrides all existing tenant configs. This would be fine if
both did client side patching, but that's not the case.
Once this merges, we must update cplane to use the PATCH endpoint.
## Summary of changes
### High Level
Allow for patching of tenant configuration with a `PATCH
/v1/tenant/config` endpoint.
It takes the same data as it's PUT counterpart. For example the payload
below will update `gc_period` and unset `compaction_period`. All other
fields are left in their original state.
```
{
"tenant_id": "1234",
"gc_period": "10s",
"compaction_period": null
}
```
### Low Level
* PS and storcon gain `PATCH /v1/tenant/config` endpoints. PS endpoint
is only used for cplane managed instances.
* `storcon_cli` is updated to have separate commands for
`set-tenant-config` and `patch-tenant-config`
Related https://github.com/neondatabase/cloud/issues/21043
## Problem
We get slru truncation commands on non-zero shards.
Compaction will drop the slru dir keys and ingest will fail when
receiving such records.
https://github.com/neondatabase/neon/pull/10080 fixed it for clog, but
not for multixact.
## Summary of changes
Only truncate multixact slrus on shard zero. I audited the rest of the
ingest code and it looks
fine from this pov.
## Problem
With pipelining enabled, the time a request spends in the batcher stage
counts towards the smgr op latency.
If pipelining is disabled, that time is not accounted for.
In practice, this results in a jump in smgr getpage latencies in various
dashboards and degrades the internal SLO.
## Solution
In a similar vein to #10042 and with a similar rationale, this PR stops
counting the time spent in batcher stage towards smgr op latency.
The smgr op latency metric is reduced to the actual execution time.
Time spent in batcher stage is tracked in a separate histogram.
I expect to remove that histogram after batching rollout is complete,
but it will be helpful in the meantime to reason about the rollout.
## Problem
In #9786 we stop storing SLRUs on non-zero shards.
However, there was one code path during ingest that still tries to
enumerate SLRU relations on all shards. This fails if it sees a tenant
who has never seen any write to an SLRU, or who has done such thorough
compaction+GC that it has dropped its SLRU directory key.
## Summary of changes
- Avoid trying to list SLRU relations on nonzero shards
## Problem
close https://github.com/neondatabase/cloud/issues/19671
```
Timeline -----------------------------
^ last GC happened LSN
^ original retention period setting = 24hr
> refresh-gc-info updates the gc_info
^ planned cutoff (gc_info)
^ customer set retention to 48hr, and it's still within the last GC LSN
^1 ^2 we have two choices: (1) update the planned cutoff to
move backwards, or (2) keep the current one
```
In this patch, we decided to keep the current cutoff instead of moving
back the gc_info to avoid races. In the future, we could allow the
planned gc cutoff to go back once cplane sends a retention_history
tenant config update, but this requires a careful revisit of the code.
## Summary of changes
Ensure that GC cutoffs never go back if retention settings get changed.
Signed-off-by: Alex Chi Z <chi@neon.tech>
Azure has a different per-request limit of 256 items for bulk deletion
compared to the number of 1000 on AWS. Therefore, we need to support
multiple values. Due to `GenericRemoteStorage`, we can't add an
associated constant, but it has to be a function.
The PR replaces the `MAX_KEYS_PER_DELETE` constant with a function of
the same name, implemented on both the `RemoteStorage` trait as well as
on `GenericRemoteStorage`.
The value serves as hint of how many objects to pass to the
`delete_objects` function.
Reading:
* https://learn.microsoft.com/en-us/rest/api/storageservices/blob-batch
* https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObjects.html
Part of #7931
## Problem
close https://github.com/neondatabase/neon/issues/10049, close
https://github.com/neondatabase/neon/issues/10030, close
https://github.com/neondatabase/neon/issues/8861
part of https://github.com/neondatabase/neon/issues/9114
The legacy gc process calls `get_latest_gc_cutoff`, which uses a Rcu
different than the gc_info struct. In the gc_compaction_smoke test case,
the "latest" cutoff could be lower than the gc_info struct, causing
gc-compaction to collect data that could be accessed by
`latest_gc_cutoff`. Technically speaking, there's nothing wrong with
gc-compaction using gc_info without considering latest_gc_cutoff,
because gc_info is the source of truth. But anyways, let's fix it.
## Summary of changes
* gc-compaction uses `latest_gc_cutoff` instead of gc_info to determine
the gc horizon.
* if a gc-compaction is scheduled via tenant compaction iteration, it
will take the gc_block lock to avoid racing with functionalities like
detach ancestor (if it's triggered via manual compaction API without
scheduling, then it won't take the lock)
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
Co-authored-by: Arpad Müller <arpad-m@users.noreply.github.com>
## Problem
In #9962 I changed the smgr metrics to include time spent on flush.
It isn't under our (=storage team's) control how long that flush takes
because the client can stop reading requests.
## Summary of changes
Stop the timer as soon as we've buffered up the response in the
`pgb_writer`.
Track flush time in a separate metric.
---------
Co-authored-by: Yuchen Liang <70461588+yliang412@users.noreply.github.com>
## Problem
part of https://github.com/neondatabase/neon/issues/9114, stacked PR
over #9809
The compaction scheduler now schedules partial compaction jobs.
## Summary of changes
* Add the compaction job splitter based on size.
* Schedule subcompactions using the compaction scheduler.
* Test subcompaction scheduler in the smoke regress test.
* Temporarily disable layer map checks
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
With the current metrics we can't identify which shards are ingesting
data at any given time.
## Summary of changes
Add a metric for the number of wal records received for processing by
each shard. This is per (tenant, timeline, shard).
We've seen cases where stray keys end up on the wrong shard. This
shouldn't happen. Add debug assertions to prevent this. In release
builds, we should be lenient in order to handle changing key ownership
policies.
Touches #9914.
## Problem
There's no metrics for disk consistent LSN and remote LSN. This stuff is
useful when looking at ingest performance.
## Summary of changes
Two per timeline metrics are added: `pageserver_disk_consistent_lsn` and
`pageserver_projected_remote_consistent_lsn`. I went for the projected
remote lsn instead of the visible one
because that more closely matches remote storage write tput. Ideally we
would have both, but these metrics are expensive.
