The `std::mem::offset_of` macro was introduced in Rust 1.77.0.
In the passing, mark the function as `const`, as suggested in the
comment. Not sure which compiler version that requires, but it works
with what have currently.
## Problem
Password hashing for sql-over-http takes up a lot of CPU. Perhaps we can
get away with temporarily caching some steps so we only need fewer
rounds, which will save some CPU time.
## Summary of changes
The output of pbkdf2 is the XOR of the outputs of each iteration round,
eg `U1 ^ U2 ^ ... U15 ^ U16 ^ U17 ^ ... ^ Un`. We cache the suffix of
the expression `U16 ^ U17 ^ ... ^ Un`. To compute the result from the
cached suffix, we only need to compute the prefix `U1 ^ U2 ^ ... U15`.
The suffix by itself is useless, which prevent's its use in brute-force
attacks should this cached memory leak.
We are also caching the full 4096 round hash in memory, which can be
used for brute-force attacks, where this suffix could be used to speed
it up. My hope/expectation is that since these will be in different
allocations, it makes any such memory exploitation much much harder.
Since the full hash cache might be invalidated while the suffix is
cached, I'm storing the timestamp of the computation as a way to
identity the match.
I also added `zeroize()` to clear the sensitive state from the
stack/heap.
For the most security conscious customers, we hope to roll out OIDC
soon, so they can disable passwords entirely.
---
The numbers for the threadpool were pretty random, but according to our
busiest region for sql-over-http, we only see about 150 unique endpoints
every minute. So storing ~100 of the most common endpoints for that
minute should be the vast majority of requests.
1 minute was chosen so we don't keep data in memory for too long.
## Problem
LKB-2502 The garbage collection of the project info cache is garbage.
What we observed: If we get unlucky, we might throw away a very hot
entry if the cache is full. The GC loop is dependent on getting a lucky
shard of the projects2ep table that clears a lot of cold entries. The GC
does not take into account active use, and the interval it runs at is
too sparse to do any good.
Can we switch to a proper cache implementation?
Complications:
1. We need to invalidate by project/account.
2. We need to expire based on `retry_delay_ms`.
## Summary of changes
1. Replace `retry_delay_ms: Duration` with `retry_at: Instant` when
deserializing.
2. Split the EndpointControls from the RoleControls into two different
caches.
3. Introduce an expiry policy based on error retry info.
4. Introduce `moka` as a dependency, replacing our `TimedLru`.
See the follow up PR for changing all TimedLru instances to use moka:
#12726.
Currently, the exporter exposes the same LFC metrics that are exposed by
the "autoscaling" sql_exporter in the docker image. With this, we can
remove the dedicated sql_exporter instance. (Actually doing the removal
is left as a TODO until this is rolled out to production and we have
changed autoscaling-agent to fetch the metrics from this new endpoint.)
The exporter runs as a Postgres background worker process. This is
extracted from the Rust communicator rewrite project, which will use the
same worker process for much more, to handle the communications with the
pageservers. For now, though, it merely handles the metrics requests.
In the future, we will add more metrics, and perhaps even APIs to
control the running Postgres instance.
The exporter listens on a Unix Domain socket within the Postgres data
directory. A Unix Domain socket is a bit unconventional, but it has some
advantages:
- Permissions are taken care of. Only processes that can access the data
directory, and therefore already have full access to the running
Postgres instance, can connect to it.
- No need to allocate and manage a new port number for the listener
It has some downsides too: it's not immediately accessible from the
outside world, and the functions to work with Unix Domain sockets are
more low-level than TCP sockets (see the symlink hack in
`postgres_metrics_client.rs`, for example).
To expose the metrics from the local Unix Domain Socket to the
autoscaling agent, introduce a new '/autoscaling_metrics' endpoint in
the compute_ctl's HTTP server. Currently it merely forwards the request
to the Postgres instance, but we could add rate limiting and access
control there in the future.
---------
Co-authored-by: Conrad Ludgate <conrad@neon.tech>
## Problem
We want to have the data-api served by the proxy directly instead of
relying on a 3rd party to run a deployment for each project/endpoint.
## Summary of changes
With the changes below, the proxy (auth-broker) becomes also a
"rest-broker", that can be thought of as a "Multi-tenant" data-api which
provides an automated REST api for all the databases in the region.
The core of the implementation (that leverages the subzero library) is
in proxy/src/serverless/rest.rs and this is the only place that has "new
logic".
---------
Co-authored-by: Ruslan Talpa <ruslan.talpa@databricks.com>
Co-authored-by: Alexander Bayandin <alexander@neon.tech>
Co-authored-by: Conrad Ludgate <conrad@neon.tech>
Second PR for the hashmap behind the updated LFC implementation ([see
first here](https://github.com/neondatabase/neon/pull/12595)). This only
adds the raw code for the hashmap/lock implementations and doesn't plug
it into the crate (that's dependent on the previous PR and should
probably be done when the full integration into the new communicator is
merged alongside `communicator-rewrite` changes?).
Some high level details: the communicator codebase expects to be able to
store references to entries within this hashmap for arbitrary periods of
time and so the hashmap cannot be allowed to move them during a rehash.
As a result, this implementation has a slightly unusual structure where
key-value pairs (and hash chains) are allocated in a separate region
with a freelist. The core hashmap structure is then an array of
"dictionary entries" that are just indexes into this region of key-value
pairs.
Concurrency support is very naive at the moment with the entire map
guarded by one big `RwLock` (which is implemented on top of a
`pthread_rwlock_t` since Rust doesn't guarantee that a
`std::sync::RwLock` is safe to use in shared memory). This (along with a
lot of other things) is being changed on the
`quantumish/lfc-resizable-map` branch.
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>
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).
## 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`.
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.
## 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`
## Problem
We need to benchmark the rich gRPC client
`client_grpc::PageserverClient` against the basic, no-frills
`page_api::Client` to determine how much overhead it adds.
Touches #11735.
Requires #12476.
## Summary of changes
Add a `pagebench --rich-client` parameter to use
`client_grpc::PageserverClient`. Also adds a compression parameter to
the client.
## Problem
The communicator gRPC client must support changing the shard map on
splits.
Touches #11735.
Requires #12476.
## Summary of changes
* Wrap the shard set in a `ArcSwap` to allow swapping it out.
* Add a new `ShardSpec` parameter struct to pass validated shard info to
the client.
* Add `update_shards()` to change the shard set. In-flight requests are
allowed to complete using the old shards.
* Restructure `get_page` to use a stable view of the shard map, and
retry errors at the top (pre-split) level to pick up shard map changes.
* Also marks `tonic::Status::Internal` as non-retryable, so that we can
use it for client-side invariant checks without continually retrying
these.
# TLDR
All changes are no-op except some metrics.
## Summary of changes I
### Pageserver
Added a new global counter metric
`pageserver_pagestream_handler_results_total` that categorizes
pagestream request results according to their outcomes:
1. Success
2. Internal errors
3. Other errors
Internal errors include:
1. Page reconstruction error: This probably indicates a pageserver
bug/corruption
2. LSN timeout error: Could indicate overload or bugs with PS's ability
to reach other components
3. Misrouted request error: Indicates bugs in the Storage Controller/HCC
Other errors include transient errors that are expected during normal
operation or errors indicating bugs with other parts of the system
(e.g., malformed requests, errors due to cancelled operations during PS
shutdown, etc.)
## Summary of changes II
This PR adds a pageserver endpoint and its counterpart in storage
controller to list visible size of all tenant shards. This will be a
prerequisite of the tenant rebalance command.
## Problem III
We need a way to download WAL
segments/layerfiles from S3 and replay WAL records. We cannot access
production S3 from our laptops directly, and we also can't transfer any
user data out of production systems for GDPR compliance, so we need
solutions.
## Summary of changes III
This PR adds a couple of tools to support the debugging
workflow in production:
1. A new `pagectl download-remote-object` command that can be used to
download remote storage objects assuming the correct access is set up.
## Summary of changes IV
This PR adds a command to list all visible delta and image layers from
index_part. This is useful to debug compaction issues as index_part
often contain a lot of covered layers due to PITR.
---------
Co-authored-by: William Huang <william.huang@databricks.com>
Co-authored-by: Chen Luo <chen.luo@databricks.com>
Co-authored-by: Vlad Lazar <vlad@neon.tech>
## Problem
The gRPC client pools don't reap idle resources.
Touches #11735.
Requires #12475.
## Summary of changes
Reap idle pool resources (channels/clients/streams) after 3 minutes of
inactivity.
Also restructure the `StreamPool` to use a mutex rather than atomics for
synchronization, for simplicity. This will be optimized later.
## Problem
The rich gRPC Pageserver client needs to split GetPage batches that
straddle multiple shards.
Touches #11735.
Requires #12462.
## Summary of changes
Adds a `GetPageSplitter` which splits `GetPageRequest` that span
multiple shards, and then reassembles the responses. Dispatches
per-shard requests in parallel.
## Problem
For the communicator, we need a rich Pageserver gRPC client.
Touches #11735.
Requires #12434.
## Summary of changes
This patch adds an initial rich Pageserver gRPC client. It supports:
* Sharded tenants across multiple Pageservers.
* Pooling of connections, clients, and streams for efficient resource
use.
* Concurrent use by many callers.
* Internal handling of GetPage bidirectional streams, with pipelining
and error handling.
* Automatic retries.
* Observability.
The client is still under development. In particular, it needs GetPage
batch splitting, shard map updates, and performance optimization. This
will be addressed in follow-up PRs.
## Problem
The communicator will need gRPC channel/client/stream pools for
efficient reuse across many backends.
Touches #11735.
Requires #12396.
## Summary of changes
Adds three nested resource pools:
* `ChannelPool` for gRPC channels (i.e. TCP connections).
* `ClientPool` for gRPC clients (i.e. `page_api::Client`). Acquires
channels from `ChannelPool`.
* `StreamPool` for gRPC GetPage streams. Acquires clients from
`ClientPool`.
These are minimal functional implementations that will need further
improvements and performance optimization. However, the overall
structure is expected to be roughly final, so reviews should focus on
that.
The pools are not yet in use, but will form the foundation of a rich
gRPC Pageserver client used by the communicator (see #12462). This PR
also adds the initial crate scaffolding for that client.
See doc comments for details.
# TLDR
All changes are no-op except
1. publishing additional metrics.
2. problem VI
## Problem I
It has come to my attention that the Neon Storage Controller doesn't
correctly update its "observed" state of tenants previously associated
with PSs that has come back up after a local data loss. It would still
think that the old tenants are still attached to page servers and won't
ask more questions. The pageserver has enough information from the
reattach request/response to tell that something is wrong, but it
doesn't do anything about it either. We need to detect this situation in
production while I work on a fix.
(I think there is just some misunderstanding about how Neon manages
their pageserver deployments which got me confused about all the
invariants.)
## Summary of changes I
Added a `pageserver_local_data_loss_suspected` gauge metric that will be
set to 1 if we detect a problematic situation from the reattch response.
The problematic situation is when the PS doesn't have any local tenants
but received a reattach response containing tenants.
We can set up an alert using this metric. The alert should be raised
whenever this metric reports non-zero number.
Also added a HTTP PUT
`http://pageserver/hadron-internal/reset_alert_gauges` API on the
pageserver that can be used to reset the gauge and the alert once we
manually rectify the situation (by restarting the HCC).
## Problem II
Azure upload is 3x slower than AWS. -> 3x slower ingestion.
The reason for the slower upload is that Azure upload in page server is
much slower => higher flush latency => higher disk consistent LSN =>
higher back pressure.
## Summary of changes II
Use Azure put_block API to uploads a 1 GB layer file in 8 blocks in
parallel.
I set the put_block block size to be 128 MB by default in azure config.
To minimize neon changes, upload function passes the layer file path to
the azure upload code through the storage metadata. This allows the
azure put block to use FileChunkStreamRead to stream read from one
partition in the file instead of loading all file data in memory and
split it into 8 128 MB chunks.
## How is this tested? II
1. rust test_real_azure tests the put_block change.
3. I deployed the change in azure dev and saw flush latency reduces from
~30 seconds to 10 seconds.
4. I also did a bunch of stress test using sqlsmith and 100 GB TPCDS
runs.
## Problem III
Currently Neon limits the compaction tasks as 3/4 * CPU cores. This
limits the overall compaction throughput and it can easily cause
head-of-the-line blocking problems when a few large tenants are
compacting.
## Summary of changes III
This PR increases the limit of compaction tasks as `BG_TASKS_PER_THREAD`
(default 4) * CPU cores. Note that `CONCURRENT_BACKGROUND_TASKS` also
limits some other tasks `logical_size_calculation` and `layer eviction`
. But compaction should be the most frequent and time-consuming task.
## Summary of changes IV
This PR adds the following PageServer metrics:
1. `pageserver_disk_usage_based_eviction_evicted_bytes_total`: captures
the total amount of bytes evicted. It's more straightforward to see the
bytes directly instead of layers.
2. `pageserver_active_storage_operations_count`: captures the active
storage operation, e.g., flush, L0 compaction, image creation etc. It's
useful to visualize these active operations to get a better idea of what
PageServers are spending cycles on in the background.
## Summary of changes V
When investigating data corruptions, it's useful to search the base
image and all WAL records of a page up to an LSN, i.e., a breakdown of
GetPage@LSN request. This PR implements this functionality with two
tools:
1. Extended `pagectl` with a new command to search the layer files for a
given key up to a given LSN from the `index_part.json` file. The output
can be used to download the files from S3 and then search the file
contents using the second tool.
Example usage:
```
cargo run --bin pagectl index-part search --tenant-id 09b99ea3239bbb3b2d883a59f087659d --timeline-id 7bedf4a6995baff7c0421ff9aebbcdab --path ~/Downloads/corruption/index_part.json-0000000c-formatted --key 000000067F000080140000802100000D61BD --lsn 70C/BF3D61D8
```
Example output:
```
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F0000801400008028000002FEFF__000007089F0B5381-0000070C7679EEB9-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000000000000000000000000000000000-000000067F0000801400008028000002F3F1__000006DD95B6F609-000006E2BA14C369-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F000080140000802100001B0973__000006D33429F539-000006DD95B6F609-0000000c
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000164D81__000006C6343B2D31-000006D33429F539-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F0000801400008021000017687B__000006BA344FA7F1-000006C6343B2D31-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000165BAB__000006AD34613D19-000006BA344FA7F1-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F0000801400000B180000000002-000000067F00008014000080210000137A39__0000069F34773461-000006AD34613D19-0000000b
tenants/09b99ea3239bbb3b2d883a59f087659d-0304/timelines/7bedf4a6995baff7c0421ff9aebbcdab/000000067F000080140000802100000D4000-000000067F000080140000802100000F0000__0000069F34773460-0000000b
```
2. Added a unit test to search the layer file contents. It's not
implemented part of `pagectl` because it depends on some test harness
code, which can only be used by unit tests.
Example usage:
```
cargo test --package pageserver --lib -- tenant::debug::test_search_key --exact --nocapture -- --tenant-id 09b99ea3239bbb3b2d883a59f087659d --timeline-id 7bedf4a6995baff7c0421ff9aebbcdab --data-dir /Users/chen.luo/Downloads/corruption --key 000000067F000080140000802100000D61BD --lsn 70C/BF3D61D8
```
Example output:
```
# omitted image for brievity
delta: 69F/769D8180: will_init: false, "OgAAALGkuwXwYp12nwYAAECGAAASIqLHAAAAAH8GAAAUgAAAIYAAAL1hDQD/DLGkuwUDAAAAEAAWAA=="
delta: 69F/769CB6D8: will_init: false, "PQAAALGkuwXotZx2nwYAABAJAAAFk7tpACAGAH8GAAAUgAAAIYAAAL1hDQD/CQUAEAASALExuwUBAAAAAA=="
```
## Problem VI
Currently when page service resolves shards from page numbers, it
doesn't fully support the case that the shard could be split in the
middle. This will lead to query failures during the tenant split for
either commit or abort cases (it's mostly for abort).
## Summary of changes VI
This PR adds retry logic in `Cache::get()` to deal with shard resolution
errors more gracefully. Specifically, it'll clear the cache and retry,
instead of failing the query immediately. It also reduces the internal
timeout to make retries faster.
The PR also fixes a very obvious bug in
`TenantManager::resolve_attached_shard` where the code tries to cache
the computed the shard number, but forgot to recompute when the shard
count is different.
---------
Co-authored-by: William Huang <william.huang@databricks.com>
Co-authored-by: Haoyu Huang <haoyu.huang@databricks.com>
Co-authored-by: Chen Luo <chen.luo@databricks.com>
Co-authored-by: Vlad Lazar <vlad.lazar@databricks.com>
Co-authored-by: Vlad Lazar <vlad@neon.tech>
See #11992 and #11961 for some examples of usecases.
This introduces a JSON serialization lib, designed for more flexibility
than serde_json offers.
## Dynamic construction
Sometimes you have dynamic values you want to serialize, that are not
already in a serde-aware model like a struct or a Vec etc. To achieve
this with serde, you need to implement a lot of different traits on a
lot of different new-types. Because of this, it's often easier to
give-in and pull all the data into a serde-aware model
(serde_json::Value or some intermediate struct), but that is often not
very efficient.
This crate allows full control over the JSON encoding without needing to
implement any extra traits. Just call the relevant functions, and it
will guarantee a correctly encoded JSON value.
## Async construction
Similar to the above, sometimes the values arrive asynchronously. Often
collecting those values in memory is more expensive than writing them as
JSON, since the overheads of `Vec` and `String` is much higher, however
there are exceptions.
Serializing to JSON all in one go is also more CPU intensive and can
cause lag spikes, whereas serializing values incrementally spreads out
the CPU load and reduces lag.
Some tests were failing with "Only request bodies with a known size
can be checksum validated." erros. This is a known issue with more
recent aws client versions, see
https://github.com/neondatabase/neon/issues/11363.
This doesn't do anything interesting yet, but demonstrates linking Rust
code to the neon Postgres extension, so that we can review and test
drive just the build process changes independently.
