duplicate MergeIterator class to avoid megamorphic calls in hot path for Row/ComplexCell and UnfilteredRowIterators
disable inlining for sinkInBinaryHeap, move typical case outside to improve hot path inlining
replace List for versions with array
fast path for same column metadata in cell merger
avoid potential megamorphic cell method invocation in ALIVE DeletionTime
minDeletionTime calculation optimization for merge logic
do not add an empty iterator, merge 2 loops into 1
patch by Dmitry Konstantinov; reviewed by Francisco Guerrero for CASSANDRA-21524
Description:
AbstractType.writeValue() is one shared method. All column types use it.
Inside writeValue(), it calls valueLengthIfFixed(). This is a virtual call. Many types override this method (Int32Type, LongType, UTF8Type, ...).
In a real cluster, many column types pass through writeValue(). So the profile always sees lots types.
Because of this, the JIT cannot inline the call. It stays as a vtable call. Also, the compiled body of writeValue() becomes big, so the JIT refuses to inline writeValue() itself ("already compiled into a big method").
We also see the same itable/vtable stubs in async-profiler output from a real cluster, running with default production options and a normal workload.
How to solve:
Add a final int field to AbstractType. Set it in the constructor. writeValue() reads this field instead of calling valueLengthIfFixed().
A field read needs no type profile. So profile pollution has no effect on it. The valueLengthIfFixed() method is not changed. Only the write path uses the field.
Result:
Production is always the polluted state, so this is the real-world comparison
JMH, JDK 17 / arm64, 1 thread, 10 x 1s warmup and measurement:
before after improvement
readValue 49.99M ops/s 56.44M ops/s +12.90%
writeValue 99.63M ops/s 114.46M ops/s +14.88%
patch by Koo Taejin; reviewed by Dmitry Konstantinov, Francisco Guerrero for CASSANDRA-21536
minDeletionTime is also added to Cell to avoid double invocation of localDeletionTime method
patch by Dmitry Konstantinov; reviewed by Francisco Guerrero for CASSANDRA-21526
A 32-bit length field read from the wire by CBUtil.readValue (and its
siblings) flowed directly into new byte[length] with no upper bound,
allowing an unauthenticated client to drive the JVM into
OutOfMemoryError: 'Requested array size exceeds VM limit' — and, with
the default -XX:OnOutOfMemoryError=kill -9 %p, terminate the process —
by declaring Integer.MAX_VALUE as the SASL-token length in AUTH_RESPONSE.
Guard the allocation in the single private readRawBytes(ByteBuf, int)
that all int32-length readers funnel through, rejecting lengths that
exceed the buffer's readable bytes with a ProtocolException.
patch by Francisco Guerrero; reviewed by Stefan Miklosovic for CASSANDRA-21521
Use a single one-time scheduled task with two tick-tock buffers to recycle metric IDs after a sufficiently long delay.
Use phantom references for ThreadLocalMetrics cleanup ony if it is needed to reduce the references processing overhead.
patch by Dmitry Konstantinov; reviewed by Benedict Elliott Smith for CASSANDRA-21475
Cassandra resolves pluggable extensions by class name from configuration, schema, and tooling
inputs. These names were loaded with an initializing Class.forName(name) and type-checked only
afterward, so the named class ran its static initializer before its type was confirmed. After this
change such classes will be loaded without initialization, verified against the expected interface or
base class, and initialized only through normal use after validation.
A shared FBUtilities.classForNameWithoutInitialization helper and typed
instanceOrConstruct/construct overloads apply this to the configurable extension points loaded by
class name: the authentication, authorization, role-management, network and
internode-authenticator backends, the partitioner, audit logger, configuration loader, seed provider,
snitch, abstract types, secondary and custom indexes, compaction strategy, compressor, replication
strategy, SASI analyzers, key and cache providers, query handler, storage and stream hooks, tracing,
the JMX authorization proxy, MBeans, the monotonic clock, nodetool Sjk, triggers, the
sstableloader and stress class options, and diagnostic event classes (loaded without initialization
and checked against DiagnosticEvent, preserving the InvalidClassException contract and the existing
package restriction).
Regression tests confirm that an invalid-type load is rejected without initializing the target
class, and that valid implementations still resolve.
Hadoop client integration and hard-coded JDK and internal class probes are left unchanged.
patch by Jeremiah Jordan; reviewed by Stefan Miklosovic for CASSANDRA-21525
SEPExecutor.maybeExecuteImmediately() runs a task synchronously on the
calling worker thread, nested within the task the worker is already
running. Such immediate tasks were invisible in system_views.queries, which only exposed each worker's primary running task.
This is common on the coordinator path, where a local read or mutation is executed immediately within the enclosing QUERY task.
Each SEPWorker now also tracks an immediate current task, set around
maybeExecuteImmediately(), and exposes it as an additional
DebuggableTaskRunner, so the queries table reports both the enclosing
task and the immediate one as separate rows.
patch by Dmitry Konstantinov; reviewed by Caleb Rackliffe for CASSANDRA-21471
Patch by Sam Tunnicliffe and Dmitry Konstantinov; reviewed by Sam
Tunnicliffe and Dmitry Konstantinov for CASSANDRA-21384
Co-authored-by: Dmitry Konstantinov <netudima@gmail.com>
Replaces the fixed-retry commit loop with deadline-based exponential
backoff for long running CMS commit operations (cms_commit_timeout=1h, 5s-60s jitter)
to allow heavily contended CMS nodes time to commit transforms.
Patch by Jon Meredith and Sam Tunnicliffe; reviewed by Jon Meredith and
Sam Tunnicliffe for CASSANDRA-21453
Co-authored-by: Sam Tunnicliffe <samt@apache.org>
- Introduced LocalCommand marker interface to identify commands that execute purely locally without a JMX connection.
- Updated History command to implement LocalCommand.
- Updated CassandraCliHelpLayout to check for LocalCommand instead of calling execution-flow methods, safely suppressing JMX options for offline commands.
patch by Arvind Kandpal; reviewed by Stefan Miklosovic, Maxim Muzafarov for CASSANDRA-20876
* Insert the actual PreInitialize entry in the distribute metadata log
table using a callback after the log is bootstrapped.
* Remove the implicit insert on subsequent commit, i.e. of the Initialize entry
* Enables the full specifics of PreInit to be encoded in the serialized form,
removing the special casing for the first CMS member and other nodes
* Correctly invalidate cached KeyspaceMetadata when in a pre-initialized state
* Update Host ids in the system.peers_v2 table directly following CMS initialization
* Initialise gossip/local host id after CMS initialization completes
* Clean up CMS initialization errors which occur after the PreInitialize stage
Co-authored-by: Alex Petrov <oleksandr.petrov@gmail.com>
Co-authored-by: Marcus Eriksson <marcuse@apache.org>
Adds an opt-in O_DIRECT write path for background SSTable producers,
bypassing the OS page cache for data that is unlikely to be re-read
soon after being written. Memtable flushes remain buffered.
Enabled via two new YAML knobs:
- background_write_disk_access_mode: standard (default) | direct
- direct_write_buffer_size: 1MiB (default; aligned up to FS block
size, auto-grown to fit a worst-case compressed chunk)
The path is gated by config, table compression being enabled, and an
OperationType allowlist in DataComponent. The allowlist is exhaustive
over OperationType: any new value left unclassified fails static
initialization.
Operations on the DIO path: COMPACTION, MAJOR_COMPACTION,
TOMBSTONE_COMPACTION, ANTICOMPACTION, GARBAGE_COLLECT, CLEANUP,
UPGRADE_SSTABLES, WRITE, STREAM (chunked receiver only), RELOCATE,
UNKNOWN (offline sstablesplit).
Operations off the DIO path:
- FLUSH (policy: just-flushed data is hot, keep in page cache)
- SCRUB (correctness: tryAppend needs mark/resetAndTruncate)
- Zero-Copy Streaming (bypasses DataComponent.buildWriter)
- Uncompressed writers (only CompressedSequentialWriter has a DIO
subclass in this change)
StartupChecks fails fast if 'direct' is requested on a platform/FS
that does not support O_DIRECT.
patch by Sam Lightfoot; reviewed by Ariel Weisberg, Dmitry Konstantinov for CASSANDRA-21134
When BTreeRow.merge reconciles an update into a row whose existing deletion
shadows the update's cells, it filtered the update (incoming) side of the merge
with Reconciler.retain, which records the removal via
PostReconciliationFunction.delete. On the memtable write path that subtracts the
shadowed cells' on-heap size from the allocator's ownership even though that
incoming data was never allocated to the memtable. Under overwrite/delete churn
that re-applies cells already covered by a newer row/partition deletion (e.g.
repair re-streaming), the allocator's "owns" counter drifts negative and the next
flush trips "AssertionError: Negative released" in MemtablePool$SubPool.released
via MemtableAllocator$SubAllocator.releaseAll during discard.
Filter the update (incoming) side with a non-recording variant,
Reconciler.removeShadowed, and keep retain() for the existing side,
whose data the memtable already owns. The filtered result is identical; only the
erroneous accounting notification is dropped. This mirrors the already-correct
complex-column path in ColumnData.merge.
patch by Stefan Miklosovic; reviewed by Dmitry Konstantinov for CASSANDRA-21469
Assisted-By: Claude Opus 4.8 <noreply@anthropic.com>
Dmitry Konstantinov