Move the `#if USE_NURAFT` guard into Context, avoiding preprocessor
directives inside `SCOPE_EXIT_SAFE` macro arguments (which is undefined
behavior). This also fixes the FreeBSD build where NuRaft is not
available.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
`tryGetKeeperDispatcher` is only available when NuRaft is enabled,
which is not the case on FreeBSD builds.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
During shutdown, `KeeperTCPHandler` connections run an infinite loop
with no shutdown detection. When the server stops the Keeper TCP
listening socket, existing handler threads keep processing heartbeats
indefinitely. `KeeperDispatcher::shutdown` (which sets the shutdown
flag) is only called after waiting for connections to close, creating
a circular dependency that blocks shutdown until the external timeout
kills the server.
Add a `shutting_down` flag to `KeeperDispatcher` that is set before
waiting for connections. `KeeperTCPHandler` checks this flag after
each `poll` return and exits cleanly when set.
https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=98447&sha=79d30b439bc793fc9aa276b28aec105750847880&name_0=PR&name_1=Stateless%20tests%20%28arm_asan%2C%20azure%2C%20parallel%29
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The previous implementation used `std::call_once` which fired in the
parent process during `BaseDaemon::initialize`. After the watchdog
`fork`, jemalloc state (e.g. `background_thread`) resets in the child
but the `once_flag` remains set, so the verification check saw a
mismatch and fired `chassert`.
Split `setup` and `verifySetup` into separate functions:
- `setup` always applies settings (idempotent), called in
`BaseDaemon::initialize` and re-applied after watchdog fork
- `verifySetup` checks jemalloc state matches server settings,
called from `Server::main` and `Keeper::main`
Also extract config key names into constants in `Jemalloc.h` to keep
them in sync between `BaseDaemon` and `ServerSettings`.
https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=98232&sha=b69293471789706abde22ec86f33104dbfaf0382&name_0=PR&name_1=Stateless%20tests%20%28amd_debug%2C%20parallel%29
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add optional <handlers> config key to type=http entries in <protocols>,
so each HTTP port can use a different set of http_handlers. This enables
use cases like running an alternative HTTP API (e.g. ES-compatible) on
a separate port with its own routing rules, without affecting the
default HTTP port.
Config example:
<protocols>
<alt_http>
<type>http</type>
<port>8124</port>
<handlers>http_handlers_alt</handlers>
</alt_http>
</protocols>
<http_handlers_alt>
<rule>...</rule>
</http_handlers_alt>
If <handlers> is omitted, the default <http_handlers> is used (fully
backward compatible). Config reload (updateServers) also respects the
per-protocol handlers key.
Netlink-based metric collection is problematic in containerized environments
(Docker, LXC, Kubernetes) where it can hang indefinitely. Benchmarks show
procfs has better tail latency and lower total CPU overhead under contention.
Since all production deployments already use the procfs path, remove netlink
entirely rather than adding a setting to disable it.
Mark getObjectStorage as const.
Add docs and tests.
Fix.
Fix drop replica.
Fix getQueryStatus.
Fix test.
Fix zookeeper usage.
Fix codes review.
Split the tests.
- Add missing `extern` for `jemalloc_profiler_sampling_rate` in
`Keeper.cpp`, `LocalServer.cpp`, and `Server.cpp`
- Pass `jemalloc_profiler_sampling_rate` as 5th argument to
`Jemalloc::setup` in `Keeper.cpp` and `LocalServer.cpp`
- Expose `prof.lg_sample` via `system.asynchronous_metrics` as
`jemalloc.prof.lg_sample`
- Add integration test `test_jemalloc_profiler_sampling_rate`
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Allow controlling jemalloc's `lg_prof_sample` via a server setting,
which determines the average interval between allocation samples.
The default value of 19 (512 KiB) matches jemalloc's built-in default.
When a non-default value is configured, `Jemalloc::setup` calls
`prof.reset` to apply the new sampling rate.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
`Server.cpp` included `StorageReplicatedMergeTree.h` but never used the
`StorageReplicatedMergeTree` class directly. The only dependency was on
`MergeTreeBackgroundExecutor` for thread pool management (accessed via
`Context::getMergeMutateExecutor` etc.).
`StorageReplicatedMergeTree.h` is an extremely heavy header that
transitively pulls in `MergeTreeData.h`, `SortCursor.h` (with expensive
`std::variant` instantiations for sorting queues), `MergeTask.h`,
`MergeFromLogEntryTask.h`, `TextIndexUtils.h`, and `DataPartsExchange.h`.
The time trace showed 3.9s just for parsing `StorageReplicatedMergeTree.h`
plus ~4.3s of template instantiations from `SortCursor.h` variant types.
Replacing it with the targeted `MergeTreeBackgroundExecutor.h` reduces
`Server.cpp.o` compile time from 34.8s to 28.1s (-19%). Since `Server.cpp`
is on the critical path (last .o file to finish before linking), this
directly reduces wall time.
Build time: 5m6.0s (was 5m16.7s, -10.7s / -3.4%)
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The config reload callback (`ConfigReloader`) fires during construction with
`initial_loading=true`. After the change to use `updateLevels(config(), ...)`
instead of `updateLevels(*config, ...)`, the callback reads `logger.level`
from the layered configuration, where the writable `MapConfiguration` at
`PRIO_APPLICATION` still holds the startup level override ("trace"). This
keeps the log level at "trace" until the explicit restoration after
`server.start()`, creating a window where HTTP requests get logged at
trace level.
Move the startup level restoration to before `server.start()` so that
no requests are ever served with the elevated startup log level.
https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=80295&sha=c7c2c8bd852fc40cf3640ee5810c7976dcfda96c&name_0=PR&name_1=Integration%20tests%20%28amd_tsan%2C%201%2F6%29
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Fix the ZK includes reload path in `Server.cpp` to use
`config().replace` instead of `removeConfiguration`/`add`/`setConfig`,
consistent with the reload callback. This ensures command-line overrides
are preserved when ZK includes are loaded.
Add integration test `test_config_not_overriding_args` that verifies
command-line arguments are not overridden by config file values
(including when ZK includes are used). Based on the test from #71452.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The `replaceByLabel` calls in `Server.cpp` and `Keeper.cpp` passed
`shared = false`, which means the `AutoPtr` inside `LayeredConfiguration`
took the raw pointer without incrementing the reference count. But the
caller's `loaded_config` `AutoPtr` also holds a reference. When the
callback's `loaded_config` goes out of scope, its destructor decrements
the reference count, potentially freeing the object while the
`LayeredConfiguration` still holds a pointer to it.
Passing `shared = true` makes the `AutoPtr` constructor call `duplicate`,
properly incrementing the reference count so the `LayeredConfiguration`
has its own counted reference.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
`BaseDaemon::initialize` performs up to 3 `chdir` calls during daemon
startup (to application path, log path, and cores directory). After
`initialize` returns, `Server::main` reads path settings from config.
If those paths are relative, they resolve against the cores directory
instead of the original startup directory.
Fix by capturing the original working directory at the start of
`BaseDaemon::initialize` and using it to resolve the main `path`
setting to absolute. All secondary path settings (like `user_files_path`,
`tmp_path`, `format_schema_path`, etc.) are resolved relative to the
now-absolute main `path`.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The config reload callback in `Server.cpp` and `Keeper.cpp` was doing
`find` + `removeConfiguration` + `add` on the shared `LayeredConfiguration`
from the `ConfigReloader` thread. This is not thread-safe because concurrent
readers (calling `getString`, etc.) iterate the same internal list without
synchronization against structural modifications.
`AbstractConfiguration` already has `mutable Poco::Mutex _mutex` that all
public read methods lock. The new `replaceByLabel` method locks the same
`_mutex` and atomically swaps the config pointer within the matching list
node — no list insert/erase needed for the common case, guaranteeing mutual
exclusion with concurrent readers.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The previous order (add new, find, remove) was incorrect because
`LayeredConfiguration::add` inserts at the same priority position
before existing entries, and `find` returns the first match.
This means the newly added config was found and removed instead
of the old one.
Changed to match the correct order used in Keeper.cpp:
find old, remove old, add new.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
