This PR simplifies our node info cache. Now we'll store entries for at
most the TTL duration, even if Redis notifications are available. This
will allow us to cache intermittent errors later (e.g. due to rate
limits) with more predictable behavior.
Related to https://github.com/neondatabase/cloud/issues/19353
## Problem
We were only resetting the limit in the wal proposer. If backends are
back pressured, it might take a while for the wal proposer to receive a
new WAL to reset the limit.
## Summary of changes
Backend also checks the time and resets the limit.
## How is this tested?
pgbench has more smooth tps
Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
Co-authored-by: Haoyu Huang <haoyu.huang@databricks.com>
## Problem
close LKB-270, close LKB-253
We periodically saw pageserver returns 404 -> storcon converts it to 500
to cplane, and causing branch operations fail. This is due to storcon is
migrating tenants across pageservers and the request was forwarded from
the storcon to pageservers while the tenant was not attached yet. Such
operations should be retried from cplane and storcon should return 503
in such cases.
## Summary of changes
- Refactor `tenant_timeline_lsn_lease` to have a single function process
and passthrough such requests: `collect_tenant_shards` for collecting
all shards and checking if they're consistent with the observed state,
`process_result_and_passthrough_errors` to convert 404 into 503 if
necessary.
- `tenant_shard_node` also checks observed state now.
Note that for passthrough shard0, we originally had a check to convert
404 to 503:
```
// Transform 404 into 503 if we raced with a migration
if resp.status() == reqwest::StatusCode::NOT_FOUND {
// Look up node again: if we migrated it will be different
let new_node = service.tenant_shard_node(tenant_shard_id).await?;
if new_node.get_id() != node.get_id() {
// Rather than retry here, send the client a 503 to prompt a retry: this matches
// the pageserver's use of 503, and all clients calling this API should retry on 503.
return Err(ApiError::ResourceUnavailable(
format!("Pageserver {node} returned 404, was migrated to {new_node}").into(),
));
}
}
```
However, this only checks the intent state. It is possible that the
migration is in progress before/after the request is processed and
intent state is always the same throughout the API call, therefore 404
not being processed by this branch.
Also, not sure about if this new code is correct or not, need second
eyes on that:
```
// As a reconciliation is in flight, we do not have the observed state yet, and therefore we assume it is always inconsistent.
Ok((node.clone(), false))
```
---------
Signed-off-by: Alex Chi Z <chi@neon.tech>
## Problem
The warning message was seen during deployment, but it's actually OK.
## Summary of changes
- Treat `"No broker updates received for a while ..."` as an info
message.
Co-authored-by: Aleksandr Sarantsev <aleksandr.sarantsev@databricks.com>
N.B: No-op for the neon-env.
## Problem
We added a per-timeline disk utilization protection circuit breaker,
which will stop the safekeeper from accepting more WAL writes if the
disk utilization by the timeline has exceeded a configured limit. We
mainly designed the mechanism as a guard against WAL upload/backup bugs,
and we assumed that as long as WAL uploads are proceeding as normal we
will not run into disk pressure. This turned out to be not true. In one
of our load tests where we have 500 PGs ingesting data at the same time,
safekeeper disk utilization started to creep up even though WAL uploads
were completely normal (we likely just maxed out our S3 upload bandwidth
from the single SK). This means the per-timeline disk utilization
protection won't be enough if too many timelines are ingesting data at
the same time.
## Summary of changes
Added a global disk utilization protection circuit breaker which will
stop a safekeeper from accepting more WAL writes if the total disk usage
on the safekeeper (across all tenants) exceeds a limit. We implemented
this circuit breaker through two parts:
1. A "global disk usage watcher" background task that runs at a
configured interval (default every minute) to see how much disk space is
being used in the safekeeper's filesystem. This background task also
performs the check against the limit and publishes the result to a
global atomic boolean flag.
2. The `hadron_check_disk_usage()` routine (in `timeline.rs`) now also
checks this global boolean flag published in the step above, and fails
the `WalAcceptor` (triggers the circuit breaker) if the flag was raised.
The disk usage limit is disabled by default.
It can be tuned with the `--max-global-disk-usage-ratio` CLI arg.
## How is this tested?
Added integration test
`test_wal_acceptor.py::test_global_disk_usage_limit`.
Also noticed that I haven't been using the `wait_until(f)` test function
correctly (the `f` passed in is supposed to raise an exception if the
condition is not met, instead of returning `False`...). Fixed it in both
circuit breaker tests.
---------
Co-authored-by: William Huang <william.huang@databricks.com>
## Problem
In the gap between picking an optimization and applying it, something
might insert a change to the intent state that makes it incompatible.
If the change is done via the `schedule()` method, we are covered by the
increased sequence number, but otherwise we can panic if we violate the
intent state invariants.
## Summary of Changes
Validate the optimization right before applying it. Since we hold the
service lock at that point, nothing else can sneak in.
Closes LKB-65
## Problem
We run multiple proxies, we get logs like
```
... spans={"http_conn#22":{"conn_id": ...
... spans={"http_conn#24":{"conn_id": ...
```
these are the same span, and the difference is confusing.
## Summary of changes
Introduce a counter per span name, rather than a global counter. If the
counter is 0, no change to the span name is made.
To follow up: see which span names are duplicated within the codebase in
different callsites
## Problem
We have several linters that use Node.js, but they are currently set up
differently, both locally and on CI.
## Summary of changes
- Add Node.js to `build-tools` image
- Move `compute/package.json` -> `build-tools/package.json` and add
`redocly` to it `@redocly/cli`
- Unify and merge into one job `lint-openapi-spec` and
`validate-compute-manifest`
## Problem
neondatabase/neon#12601 did't compleatly disable writing `*.profraw`
files, but instead of `/tmp/coverage` it started to write into the
current directory
## Summary of changes
- Set `LLVM_PROFILE_FILE=/dev/null` to avoing writing `*.profraw` at all
Initial PR for the hashmap behind the updated LFC implementation. This
refactors `neon-shmem` so that the actual shared memory utilities are in
a separate module within the crate. Beyond that, it slightly changes
some of the docstrings so that they play nicer with `cargo doc`.
We don't want to depend on postgres_ffi in an API crate. If there is no
such dependency, we can compile stuff like `storcon_cli` without needing
a full working postgres build. Fixes regression of #12548 (before we
could compile it).
- Remove a few obsolete "allowed error messages" from tests. The
pageserver doesn't emit those messages anymore.
- Remove misplaced and outdated docstring comment from
`test_tenants.py`. A docstring is supposed to be the first thing in a
function, but we had added some code before it. And it was outdated, as
we haven't supported running without safekeepers for a long time.
- Fix misc typos in comments
- Remove obsolete comment about backwards compatibility with safekeepers
without `TIMELINE_STATUS` API. All safekeepers have it by now.
## Problem
We don't use code coverage produced by `regress-tests`
(neondatabase/neon#6798), so there's no need to collect it. Potentially,
disabling it should reduce the load on disks and improve the stability
of debug builds.
## Summary of changes
- Disable code coverage collection for regression tests
## Problem
This is a prerequisite for neondatabase/neon#12575 to keep all things
relevant to `build-tools` image in a single directory
## Summary of changes
- Rename `build_tools/` to `build-tools/`
- Move `build-tools.Dockerfile` to `build-tools/Dockerfile`
## Problem
Not all cplane errors are properly recognized and cached/retried.
## Summary of changes
Add more cplane error reasons. Also, use retry_delay_ms as cache TTL if
present.
Related to https://github.com/neondatabase/cloud/issues/19353
All Errors that can occur during get_installed_extensions() come from
tokio-postgres functions, e.g. if the database is being shut down
("FATAL: terminating connection due to administrator command"). I'm
seeing a lot of such errors in the logs with the regression tests, with
very verbose stack traces. The compute_ctl stack trace is pretty useless
for errors originating from the Postgres connection, the error message
has all the information, so stop printing the stack trace.
I changed the result type of the functions to return the originating
tokio_postgres Error rather than anyhow::Error, so that if we introduce
other error sources to the functions where the stack trace might be
useful, we'll be forced to revisit this, probably by introducing a new
Error type that separates postgres errors from other errors. But this
will do for now.
We didn't consistently apply these, and it wasn't consistently solved.
With this patch we should have a more consistent approach to this, and
have less issues porting changes to newer versions.
This also removes some potentially buggy casts to `long` from `uint64` -
they could've truncated the value in systems where `long` only has 32
bits.
Include records and image in the debug get page handler.
This endpoint does not update the metrics and does not support tracing.
Note that this now returns individual bytes which need to be encoded
properly for debugging.
Co-authored-by: Haoyu Huang <haoyu.huang@databricks.com>
# TLDR
This PR is a no-op.
## Problem
When a SK loses a disk, it must recover all WALs from the very
beginning. This may take days/weeks to catch up to the latest WALs for
all timelines it owns.
## Summary of changes
When SK starts up,
if it finds that it has 0 timelines,
- it will ask SC for the timeline it owns.
- Then, pulls the timeline from its peer safekeepers to restore the WAL
redundancy right away.
After pulling timeline is complete, it will become active and accepts
new WALs.
The current impl is a prototype. We can optimize the impl further, e.g.,
parallel pull timelines.
---------
Co-authored-by: Haoyu Huang <haoyu.huang@databricks.com>
https://github.com/neondatabase/cloud/issues/19011
Measure relative performance for prewarmed and non-prewarmed endpoints.
Add test that runs on every commit, and one performance test with a
remote cluster.
## Problem
In https://github.com/neondatabase/neon/pull/12513, the new code was
implemented to retry 404 errors caused by the replication lag. However,
this implemented the new logic, making the script more complicated,
while we have an existing one in `neon_api.py`.
## Summary of changes
The existing mechanism is used to retry 404 errors.
---------
Co-authored-by: Alexey Masterov <alexey.masterov@databricks.com>
## Problem
The communicator gRPC client currently uses bounded client/stream pools.
This can artificially constrain clients, especially after we remove
pipelining in #12584.
[Benchmarks](https://github.com/neondatabase/neon/pull/12583) show that
the cost of an idle server-side GetPage worker task is about 26 KB (2.5
GB for 100,000), so we can afford to scale out.
In the worst case, we'll degenerate to the current libpq state with one
stream per backend, but without the TCP connection overhead. In the
common case we expect significantly lower stream counts due to stream
sharing, driven e.g. by idle backends, LFC hits, read coalescing,
sharding (backends typically only talk to one shard at a time), etc.
Currently, Pageservers rarely serve more than 4000 backend connections,
so we have at least 2 orders of magnitude of headroom.
Touches #11735.
Requires #12584.
## Summary of changes
Remove the pool limits, and restructure the pools.
We still keep a separate bulk pool for Getpage batches of >4 pages (>32
KB), with fewer streams per connection. This reduces TCP-level
congestion and head-of-line blocking for non-bulk requests, and
concentrates larger window sizes on a smaller set of
streams/connections, presumably reducing memory usage. Apart from this,
bulk requests don't have any latency penalty compared to other requests.
## Problem
We don't log the timeline ID when rolling ephemeral layers during
housekeeping.
Resolves [LKB-179](https://databricks.atlassian.net/browse/LKB-179)
## Summary of changes
Add a span with timeline ID when calling `maybe_freeze_ephemeral_layer`
from the housekeeping loop.
We don't instrument the function itself, since future callers may not
have a span including the tenant_id already, but we don't want to
duplicate the tenant_id for these spans.
## Problem
The communicator gRPC client currently attempts to pipeline GetPage
requests from multiple callers onto the same gRPC stream. This has a
number of issues:
* Head-of-line blocking: the request may block on e.g. layer download or
LSN wait, delaying the next request.
* Cancellation: we can't easily cancel in-progress requests (e.g. due to
timeout or backend termination), so it may keep blocking the next
request (even its own retry).
* Complex stream scheduling: picking a stream becomes harder/slower, and
additional Tokio tasks and synchronization is needed for stream
management.
Touches #11735.
Requires #12579.
## Summary of changes
This patch removes pipelining of gRPC stream requests, and instead
prefers to scale out the number of streams to achieve the same
throughput. Stream scheduling has been rewritten, and mostly follows the
same pattern as the client pool with exclusive acquisition by a single
caller.
[Benchmarks](https://github.com/neondatabase/neon/pull/12583) show that
the cost of an idle server-side GetPage worker task is about 26 KB (2.5
GB for 100,000), so we can afford to scale out.
This has a number of advantages:
* It (mostly) eliminates head-of-line blocking (except at the TCP
level).
* Cancellation becomes trivial, by closing the stream.
* Stream scheduling becomes significantly simpler and cheaper.
* Individual callers can still use client-side batching for pipelining.
## Problem
The new communicator gRPC client has significantly worse Pagebench
performance than a basic gRPC client. We need to find out why.
## Summary of changes
Add a `pagebench --profile` flag which takes a client CPU profile of the
benchmark and writes a flamegraph to `profile.svg`.
## Problem
It can take 3x the idle timeout to reap a channel. We have to wait for
the idle timeout to trigger first for the stream, then the client, then
the channel.
Touches #11735.
## Summary of changes
Reap empty channels immediately, and rely indirectly on the
channel/stream timeouts.
This can still lead to 2x the idle timeout for streams (first stream
then client), but that's okay -- if the stream closes abruptly (e.g. due
to timeout or error) we want to keep the client around in the pool for a
while.
## Problem
gRPC client retries currently include pool acquisition under the
per-attempt timeout. If pool acquisition is slow (e.g. full pool), this
will cause spurious timeout warnings, and the caller will lose its place
in the pool queue.
Touches #11735.
## Summary of changes
Makes several improvements to retries and related logic:
* Don't include pool acquisition time under request timeouts.
* Move attempt timeouts out of `Retry` and into the closure.
* Make `Retry` configurable, move constants into main module.
* Don't backoff on the first retry, and reduce initial/max backoffs to
5ms and 5s respectively.
* Add `with_retries` and `with_timeout` helpers.
* Add slow logging for pool acquisition, and a `warn_slow` counterpart
to `log_slow`.
* Add debug logging for requests and responses at the client boundary.
## Problem
For the communicator scheduling policy, we need to understand the
server-side cost of idle gRPC streams.
Touches #11735.
## Summary of changes
Add an `idle-streams` benchmark to `pagebench` which opens a large
number of idle gRPC GetPage streams.
## Problem
When refreshing cancellation data we resend the entire value again just
to reset the TTL, which causes unnecessary load in proxy, on network and
possibly on redis side.
## Summary of changes
* Switch from using SET with full value to using EXPIRE to reset TTL.
* Add a tiny delay between retries to prevent busy loop.
* Shorten CancelKeyOp variants: drop redundant suffix.
* Retry SET when EXPIRE failed.
## Problem
When a connection terminates its maintain_cancel_key task keeps running
until the CANCEL_KEY_REFRESH sleep finishes and then it triggers another
cancel key TTL refresh before exiting.
## Summary of changes
* Check for cancellation while sleeping and interrupt sleep.
* If cancelled, break the loop, don't send a refresh cmd.
## Problem
We don't validate the validity of the `new_sk_set` before starting the
migration. It is validated later, so the migration to an invalid
safekeeper set will fail anyway. But at this point we might already
commited an invalid `new_sk_set` to the database and there is no `abort`
command yet (I ran into this issue in neon_local and ruined the timeline
:)
- Part of https://github.com/neondatabase/neon/issues/11669
## Summary of changes
- Add safekeeper count and safekeeper duplication checks before starting
the migration
- Test that we validate the `new_sk_set` before starting the migration
- Add `force` option to the `TimelineSafekeeperMigrateRequest` to
disable not-mandatory checks
Serialize query row responses directly into JSON. Some of this code
should be using the `json::value_as_object/list` macros, but I've
avoided it for now to minimize the size of the diff.
## Problem
To store cancellation data we send two commands to redis because the
redis server version doesn't support HSET with EX. Also, HSET is not
really needed.
## Summary of changes
* Replace the HSET + EXPIRE command pair with one SET .. EX command.
* Replace HGET with GET.
* Leave a workaround for old keys set with HSET.
* Replace some anyhow errors with specific errors to surface the
WRONGTYPE error from redis.
# TLDR
Problem-I is a bug fix. The rest are no-ops.
## Problem I
Page server checks image layer creation based on the elapsed time but
this check depends on the current logical size, which is only computed
on shard 0. Thus, for non-0 shards, the check will be ineffective and
image creation will never be done for idle tenants.
## Summary of changes I
This PR fixes the problem by simply removing the dependency on current
logical size.
## Summary of changes II
This PR adds a timeout when calling page server to split shard to make
sure SC does not wait for the API call forever. Currently the PR doesn't
adds any retry logic because it's not clear whether page server shard
split can be safely retried if the existing operation is still ongoing
or left the storage in a bad state. Thus it's better to abort the whole
operation and restart.
## Problem III
`test_remote_failures` requires PS to be compiled in the testing mode.
For PS in dev/staging, they are compiled without this mode.
## Summary of changes III
Remove the restriction and also increase the number of total failures
allowed.
## Summary of changes IV
remove test on PS getpage http route.
---------
Co-authored-by: Chen Luo <chen.luo@databricks.com>
Co-authored-by: Yecheng Yang <carlton.yang@databricks.com>
Co-authored-by: Vlad Lazar <vlad@neon.tech>
## Problem
close LKB-162
close https://github.com/neondatabase/cloud/issues/30665, related to
https://github.com/neondatabase/cloud/issues/29434
We see a lot of errors like:
```
2025-05-22T23:06:14.928959Z ERROR compaction_loop{tenant_id=? shard_id=0304}:run:gc_compact_timeline{timeline_id=?}: error applying 4 WAL records 35/DC0DF0B8..3B/E43188C0 (8119 bytes) to key 000000067F0000400500006027000000B9D0, from base image with LSN 0/0 to reconstruct page image at LSN 61/150B9B20 n_attempts=0: apply_wal_records
Caused by:
0: read walredo stdout
1: early eof
```
which is an acceptable form of error and we should downgrade it to
warning.
## Summary of changes
walredo error during gc-compaction is expected when the data below the
gc horizon does not contain a full key history. This is possible in some
rare cases of gc that is only able to remove data in the middle of the
history but not all earlier history when a full keyspace gets deleted.
Signed-off-by: Alex Chi Z <chi@neon.tech>
Fixes [LKB-61](https://databricks.atlassian.net/browse/LKB-61):
`test_timeline_archival_chaos` being flaky with storcon error `Requested
tenant is missing`.
When a tenant migration is ongoing, and the attach request has been sent
to the new location, but the attach hasn't finished yet, it is possible
for the pageserver to return a 412 precondition failed HTTP error on
timeline deletion, because it is being sent to the new location already.
That one we would previously log via sth like:
```
ERROR request{method=DELETE path=/v1/tenant/1f544a11c90d1afd7af9b26e48985a4e/timeline/32818fb3ebf07cb7f06805429d7dee38 request_id=c493c04b-7f33-46d2-8a65-aac8a5516055}: Error processing HTTP request: InternalServerError(Error deleting timeline 32
818fb3ebf07cb7f06805429d7dee38 on 1f544a11c90d1afd7af9b26e48985a4e on node 2 (localhost): pageserver API: Precondition failed: Requested tenant is missing
```
This patch changes that and makes us return a more reasonable resource
unavailable error. Not sure how scalable this is with tenants with a
large number of shards, but that's a different discussion (we'd probably
need a limited amount of per-storcon retries).
example
[link](https://neon-github-public-dev.s3.amazonaws.com/reports/pr-12398/15981821532/index.html#/testresult/e7785dfb1238d92f).
Update the WSS estimate before acquring the lock, so that we don't need
to hold the lock for so long. That seems safe to me, see added comment.
I was planning to do this with the new rust-based communicator
implementation anyway, but it might help a little with the current C
implementation too. And more importantly, having this as a separate PR
gives us a chance to review this aspect independently.
## Problem
Canceelation requires redis, redis required control-plane.
## Summary of changes
Make redis for cancellation not require control plane.
Add instructions for setting up redis locally.
Split the functions into two: one internal function to calculate the
estimate, and another (two functions) to expose it as SQL functions.
This is in preparation of adding new communicator implementation. With
that, the SQL functions will dispatch the call to the old or new
implementation depending on which is being used.
This is a no-op for the neon deployment
* Introduce the concept image consistent lsn: of the largest LSN below
which all pages have been redone successfully
* Use the image consistent LSN for forced image layer creations
* Optionally expose the image consistent LSN via the timeline describe
HTTP endpoint
* Add a sharded timeline describe endpoint to storcon
---------
Co-authored-by: Chen Luo <chen.luo@databricks.com>
## Problem
We have a `safekeeper_migrate` handler, but no subcommand in
`storcon_cli`. Same for `/:timeline_id/locate` for identifying current
set of safekeepers.
- Closes: https://github.com/neondatabase/neon/issues/12395
## Summary of changes
- Add `timeline-safekeeper-migrate` and `timeline-locate` subcommands to
`storcon_cli`
On December 8th, 2023, an engineering escalation (INC-110) was opened
after it was found that BYPASSRLS was being applied to all roles.
PR that introduced the issue:
https://github.com/neondatabase/neon/pull/5657
Subsequent commit on main:
ad99fa5f03
NOBYPASSRLS and INHERIT are the defaults for a Postgres role, but
because it isn't easy to know if a Postgres cluster is affected by the
issue, we need to keep the migration around for a long time, if not
indefinitely, so any cluster can be fixed.
Branching is the gift that keeps on giving...
Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
