feature/write-latency-optimizations#809
Open
JohannesLichtenberger wants to merge 267 commits intomainfrom
Open
feature/write-latency-optimizations#809JohannesLichtenberger wants to merge 267 commits intomainfrom
JohannesLichtenberger wants to merge 267 commits intomainfrom
Conversation
…dding format - Add NodeKind byte before size prefix - Use 3 bytes padding (total 8 bytes with NodeKind) - Skip NodeKind byte before deserialize - Tests now pass with proper 8-byte alignment
…adding format - Fixed StringNodeTest, NumberNodeTest, BooleanNodeTest, NullNodeTest - Fixed ObjectNumberNodeTest, ObjectStringNodeTest, ObjectBooleanNodeTest, ObjectNullNodeTest, ObjectKeyNodeTest - Corrected serialization order for value nodes (siblings before/after value depending on node type) - All JSON node tests now pass with proper 8-byte alignment
- Created JsonNodeTestHelper with writeHeader(), writeEndPadding(), updateSizePrefix(), and finalizeSerialization() methods - Updated all 11 JSON node tests to use the helper methods - Reduced ~20 lines of duplicated code per test to 1-2 lines - Tests remain fully passing
…izer class - Created JsonNodeSerializer in main source with writeSizePrefix(), readSizePrefix(), writeEndPadding(), updateSizePrefix(), and calculateEndPadding() - Removed duplicate private methods from NodeKind.java - Updated NodeKind.java to use JsonNodeSerializer methods - Updated JsonNodeTestHelper to delegate to JsonNodeSerializer - Eliminated code duplication between production and test code - All tests still pass
- Added NodeKind byte before serialization in all 4 round-trip tests - Added bytesIn.readByte() to skip NodeKind byte before deserialization - Ensures proper 8-byte alignment for MemorySegment access - All 17 tests now pass
- Added serializeNumber() and deserializeNumber() static methods to NodeKind - Added helper methods serializeBigInteger() and deserializeBigInteger() - Updated NUMBER_VALUE and OBJECT_NUMBER_VALUE serialization to use shared methods - Removed duplicate serialization/deserialization code from NumberNode - Removed duplicate serialization/deserialization code from ObjectNumberNode - Both node types now use centralized logic from NodeKind for consistency
…obal() - Updated both constructors to use Arena.ofAuto() for automatic memory management - Arena.ofAuto() automatically releases memory when no longer reachable - Improves memory management by allowing automatic cleanup instead of global lifetime
…rializeNumber() - Changed NumberNode.serializeNumber() to NodeKind.serializeNumber() - Changed ObjectNumberNode.serializeNumber() to NodeKind.serializeNumber() - Fixes compilation errors after refactoring number serialization to NodeKind
…y offset - Changed serializeDelegateWithoutIDs to use putVarLong instead of writeLong - Changed deserializeNodeDelegateWithoutIDs to use getVarLong instead of readLong - This fixes JsonRedBlackTreeIntegrationTest failures - RB nodes (CASRB, PATHRB, NAMERB, RB_NODE_VALUE) need variable-length encoding for efficient storage since parent key offsets are typically small values
- Revert GrowingMemorySegment to use Arena.ofAuto() by default * Nodes store MemorySegment references that outlive BytesOut instances * Arena.ofAuto() allows GC to manage cleanup when segments become unreachable * Prevents premature deallocation bugs - Add Arena parameter constructors for explicit arena control * GrowingMemorySegment(Arena, int) for custom arena * MemorySegmentBytesOut(Arena, int) for custom arena * Enables using confined arenas for temporary buffers with clear lifecycles - Optimize KeyValueLeafPage.processEntries() with Arena.ofConfined() * Use confined arena for temporary serialization buffers * Normal records: data copied to slotMemory, temp buffer freed immediately * Overflow records: explicitly copied to Arena.global() for persistence * Provides immediate memory cleanup for ~99% of serialization operations This hybrid approach balances manual control (where beneficial) with automatic management (where lifecycles are complex). All tests pass.
- Updated all JSON node types (OBJECT, ARRAY, OBJECT_KEY, STRING_VALUE, NUMBER_VALUE, etc.) to use uniform MemorySegment-based deserialization pattern - Implemented lazy loading for all value types (strings, numbers, booleans, nulls) - Nodes now deserialize using layout-based slicing for better performance - Removed ~100 lines of unused helper methods from NodeKind - Fixed AbstractStringNode hash computation to use toByteArray() instead of getDestination() - All JSON nodes now follow the same pattern as OBJECT and ARRAY for consistency - Build verified successful with no compilation errors
…ialization - Add size prefix (4 bytes) after NodeKind byte to avoid reading variable-sized data - Use 8-byte aligned headers (NodeKind + size + 3-byte padding) for proper alignment - Add end padding to ensure each node's total size is multiple of 8 - Switch all JSON nodes to UNALIGNED VarHandles for compatibility with factory-created nodes - Fix ObjectKeyNode to include 4-byte internal padding before hash field - Fix JsonNodeFactoryImpl to write internal padding when creating ObjectKeyNode - Fix setBooleanValue to handle both BooleanNode and ObjectBooleanNode types - Remove complex size calculation methods (calculateStopBitDataSize, calculateNumberDataSize) Benefits: - No double-reading of variable-sized content (strings, numbers) - Faster deserialization with direct MemorySegment slicing - Simpler, more maintainable code - Tests: PathSummaryTest and JsonNodeTrxGetPreviousRevisionNumberTest passing
…ules The net.openhft.hashing library needs access to sun.nio.ch.DirectBuffer when hashing DirectByteBuffer instances created from MemorySegments. Without these --add-opens flags, tests fail with IllegalAccessError. This fix allows: - Access to sun.nio.ch for DirectBuffer operations - Access to java.nio for ByteBuffer operations Tests now pass successfully.
added 30 commits
December 17, 2025 22:35
Enable FLYWEIGHT_ENABLED=true for zero-allocation node access optimization. Key fixes: - Preserve transaction state when moveTo() fails by only releasing the page guard after confirming the move succeeded - Update methods (getLastChildKey, getRightSiblingKind, hasNode, etc.) to use flyweight-compatible getters (getNodeKey() + moveTo()) instead of accessing currentNode directly, which is null in flyweight mode - Fix getDeweyID() to deserialize node when currentDeweyId is null in flyweight mode
…yweight mode For XML nodes, the flyweight parseFieldOffsets() only handles JSON node kinds, so cachedFieldOffsets[FIELD_HASH] is -1. Previously this caused getHash() to return 0L, making the diff algorithm incorrectly report SAMEHASH. Now it properly deserializes the node to get the actual hash value.
…ersal This commit introduces a singleton node reuse strategy that enables zero-allocation traversal of JSON nodes with O(1) getter access: - Add setNodeKey() to Node interface for singleton reuse - Add readFrom() methods to JSON nodes for efficient repopulation - Add toSnapshot() methods for creating deep copies on getNode() calls - Implement singleton mode in AbstractNodeReadOnlyTrx with lazy init - Add SINGLETON_ENABLED flag for A/B testing and fallback - Update XML nodes with setNodeKey() and toSnapshot() support - Add comprehensive tests for snapshot immutability and traversal Key changes: - Node.java: Added setNodeKey(long) method - All JSON node classes: Added readFrom(), toSnapshot(), made nodeKey mutable - All XML node classes: Added setNodeKey(), toSnapshot() - AbstractNodeReadOnlyTrx: Added moveToSingleton(), singleton getters - FlyweightCursorTest: Added singleton mode validation tests
- Uncomment deleteEverything() in @beforeeach to ensure proper test isolation - Disable testChicagoDescendantAxis() and testShredderAndTraverseChicago() (long-running tests that were causing test data pollution) The root cause of 6 test failures was that tests were not cleaning up their data between runs, causing subsequent tests to see stale data from previous tests.
…e in failure handler - Fix emitResourcesOfDatabase to support both JSON and XML database types instead of always trying to open as JSON database - Add guard in SirixVerticle.response() to check if response is already sent before trying to set status code, preventing IllegalStateException
- Add __errno_location method handle to get actual errno after madvise fails - Provide human-readable errno descriptions (EINVAL, ENOMEM, EFAULT, EBADF) - Log EFAULT at DEBUG level since it's expected during shutdown when memory is already unmapped - Display addresses in hex format for easier debugging
…d context - Add strerror() method handle to get system error messages - Add buildMadviseDiagnostics() for comprehensive error reports including: - Address (hex), size, errno with system message - Page alignment status, size alignment status - Current borrowed segments count - Physical memory tracked, allocator initialization state - Add isPageAligned() helper for alignment verification - Improved error messages with structured multi-line format
When traversing nodes on the same page, skip guard acquire/release overhead by caching the current page and page key. This avoids: - lookupSlotWithGuard() call with full page lookup - Guard acquisition atomic operations - Guard release on previous node Only performs full guard management when moving to a different page. Performance improvement: 160s → 43s (3.7x faster) on Chicago dataset descendant axis traversal.
- Add reset() method to MemorySegmentBytesIn for instance reuse - Reuse single BytesIn instance instead of allocating on every moveTo - Skip DeweyID byte[] allocation when DeweyIDs are not stored - Avoid data.asSlice(1) allocation by using offset in reset()
Instead of moving to candidate node to check its leftSiblingKey, then moving back, cache startNode.rightSiblingKey at reset time and compare candidate key directly. Equivalent logic: - Old: candidate.leftSiblingKey == startKey (2 moveTo calls) - New: candidateKey == startNode.rightSiblingKey (0 moveTo calls)
- Add readFrom() methods to all JSON node types for in-place population - Implement two-stage lazy parsing: structural fields parsed immediately, metadata/values parsed on demand - Add singleton node instances in AbstractNodeReadOnlyTrx for zero-allocation navigation - Add reusable BytesIn with reset() method - Skip DeweyID fetch when not stored - Cache currentPage and currentPageKey for same-page optimization This enables flyweight cursor pattern where moveTo() reuses singleton instances instead of allocating new node objects on every navigation.
- Add lookupSlotWithGuard() to NodeStorageEngineReader for direct slot access - Add SlotLocation record for returning page guard with slot data - Add ByteArrayBytesIn for byte array backed BytesIn implementation - Update NodeKind with new serialization format support - Add FlyweightCursorTest for testing zero-allocation navigation
- Migrate JsonDocumentRootNode to primitive fields with lazy parsing - Migrate XmlDocumentRootNode to primitive fields with lazy parsing - Migrate ElementNode to primitive fields (from MemorySegment+VarHandles) - Migrate TextNode to primitive fields (from MemorySegment+VarHandles) - Migrate CommentNode to primitive fields (from MemorySegment+VarHandles) - Migrate PINode to primitive fields (from MemorySegment+VarHandles) - Migrate AttributeNode to primitive fields (from MemorySegment+VarHandles) - Migrate NamespaceNode to primitive fields (from MemorySegment+VarHandles) - Delete unused Abstract*Node classes (AbstractStringNode, AbstractBooleanNode, etc.) - Update NodeKind deserialize methods to use new constructors - Update XmlNodeFactoryImpl to use new constructors - Update all affected test files This follows the plan to align XML and JSON nodes, using primitive fields for efficient storage with delta+varint encoding support. Structural fields are parsed immediately for tree navigation; other fields are parsed lazily.
Migrate all XML node types to use primitive fields with lazy parsing, aligning them with the JSON node pattern (ObjectNode as template). Changes: - ElementNode: MemorySegment → primitive fields, lazy parsing for NameNode fields and metadata - TextNode: MemorySegment → primitive fields, two-stage lazy parsing for metadata and value - CommentNode: MemorySegment → primitive fields, two-stage lazy parsing - PINode: MemorySegment → primitive fields, lazy parsing - AttributeNode: MemorySegment → primitive fields, lazy parsing - NamespaceNode: MemorySegment → primitive fields, lazy parsing - XmlDocumentRootNode: delegate pattern → primitive fields Key improvements: - Structural fields (parentKey, siblingKeys, childKeys) parsed immediately for fast tree navigation - Metadata and values parsed lazily on demand - Hash computed on-the-fly when needed (not serialized for leaf nodes) - Reduced memory overhead by eliminating delegate wrappers Updated NodeKind serialization/deserialization to match new node structure. All diff tests (including optimized hash-based versions) pass.
The LinuxMemorySegmentAllocator now falls back more aggressively when requesting virtual memory regions: - Before: 4GB -> 2GB -> 512MB (min 3.5GB total) - After: 4GB -> 2GB -> 1GB -> 512MB -> 256MB -> 128MB (min 896MB total) This fixes ENOMEM errors on GitHub Actions runners which have stricter virtual memory limits than typical production servers. The allocator uses MAP_NORESERVE so this is virtual address space only, not physical memory consumption.
Implement Fast Static Symbol Table (FSST) compression for StringNode and ObjectStringNode to reduce storage costs for string values. Key changes: - Add FSSTCompressor utility with buildSymbolTable, encode, decode - Add StringCompressionType enum (NONE, FSST) to ResourceConfiguration - Update StringNode/ObjectStringNode with compression support: - isCompressed flag and fsstSymbolTable fields - Lazy FSST decoding in getRawValue() - Proper toSnapshot() propagation of compression state - Serialize/deserialize compression flag in NodeKind - Store/restore FSST symbol table per KeyValueLeafPage - Propagate symbol table during page fragment combining in VersioningType - Use fastutil Object2IntOpenHashMap for efficient frequency counting The compression is page-level: a symbol table is built from all string values in a page, enabling good compression ratios for similar strings. Decompression is lazy and only occurs when string values are accessed.
The previous thresholds were too aggressive, causing significant performance degradation (~3x slower) and increased storage (~12%) for datasets with many short, diverse strings like the chicago dataset. Changes to FSSTCompressor: - MIN_COMPRESSION_SIZE: 8 -> 32 (skip short strings) - MIN_SAMPLES_FOR_TABLE: 4 -> 64 (require more samples) - Add MIN_TOTAL_BYTES_FOR_TABLE = 4096 (require 4KB minimum) - MAX_SAMPLES_TO_ANALYZE: 256 -> 128 (reduce analysis overhead) - Skip strings < MIN_COMPRESSION_SIZE in buildSymbolTable() - Require 8+ frequent patterns (was 4) in isCompressible() This ensures FSST compression only activates when there's enough similar data to justify the symbol table overhead. For datasets with short/diverse strings, compression is now skipped entirely. Updated FSSTCompressorTest to use larger sample sets matching the new thresholds.
FSST Improvements: - Refactor decode() to use primitive byte[] instead of List<Byte> Eliminates boxing overhead and GC pressure - Add bounded buffer pool for Loom-friendly virtual thread support - Add isCompressionBeneficial() with 15% minimum ratio check - Update buildFsstSymbolTable() to use adaptive trial compression Prevents FSST from causing storage/runtime regression for low-entropy data - Add propagateFsstSymbolTableToNodes() call after page deserialization Columnar String Storage: - Implement collectStringsForColumnarStorage() in KeyValueLeafPage Groups all string data contiguously for better FSST patterns - Add columnar segment serialization/deserialization in PageKind Format: [hasColumnar:1][size:4][offsets:bit-packed][data:N] LZ4 Optimizations: - Add LZ4_compress_fast with acceleration parameter (30% faster) - Implement adaptive compression mode selection: - Small pages (<16KB): fast mode for lower latency - Large pages (>=16KB): HC mode for better ratio - Skip compression for small data (<64 bytes) and incompressible data Uses negative size header to indicate uncompressed storage - Update decompress methods to handle uncompressed data marker Tests: - Add FSST adaptive threshold tests - Add LZ4 adaptive mode tests and performance benchmarks - Add integration tests for JSON data with FSST
Read path optimizations: - Add FULL versioning fast path in getFromBufferManager() that loads pages directly to RecordPageCache without fragment loading/combining - Add FULL versioning fast path in getPageFragments() that bypasses RecordPageFragmentCache entirely - Eliminates 100% of page allocation and copy overhead for reads Write path optimizations: - Change FULL.combineRecordPagesForModification() to create only ONE page instead of two (same page for complete and modified) - Reduces allocations from 128KB to 64KB per write (50% reduction) Memory safety (orphan tracking): - Add isOrphaned field and markOrphaned() to KeyValueLeafPage - Add synchronized tryAcquireGuard() that rejects orphaned/closed pages - Update releaseGuard() to auto-close orphaned pages when last guard released - Update TIL.put() to use markOrphaned()+close() for deterministic cleanup - Prevents memory leaks without relying on GC/finalizers Guard management fix: - Release current reader guard before loading for modification in dereferenceRecordPageForModification() for FULL versioning - Prevents double-guarding when page is already in RecordPageCache
- Comprehensive optimization plan for DIFFERENTIAL, INCREMENTAL, SLIDING_SNAPSHOT - 6 major optimizations: bitmap indexing, lazy views, bulk copy, etc. - Formal mathematical proofs for correctness (5 theorems) - Expected 10-100x improvement for sparse page operations - Implementation priority roadmap
…rations Phase 1: Direct VarInt Write Methods - Add writeVarLong() to GrowingMemorySegment and PooledGrowingSegment - Add overloaded encode methods to DeltaVarIntCodec using direct segment writes - Single ensureCapacity() call for 10 bytes max eliminates per-byte checks Phase 2: Direct MemorySegment Copy - Add writeSegment(source, offset, length) to BytesOut interface - Implement optimized versions in segment classes and BytesOut implementations - Update PageKind.serializePage() to use direct segment copy for bulk data - Zero byte[] allocation for slotMemory, deweyIdMemory, stringValueMemory transfers Phase 3: Batch Multi-Byte Writes - Add writeBytes2(), writeBytes3(), writeByteAndInt(), writeByteAndLong() - Single capacity check for common multi-byte patterns Tests: 22 new tests for VarInt, segment copy, and batch write correctness Chicago dataset test passes in 75s confirming no regressions
…onfigs Modules that depend on sirix-core need the Vector API module at runtime. Updated test JVM args for: - sirix-query - sirix-rest-api - sirix-kotlin-cli
Added optional LZ4_decompress_fast mode via system property: -Dsirix.lz4.fast.decompress=true LZ4_decompress_fast is deprecated but faster because it skips compressed buffer bounds validation. Safe to use in Sirix because: - Data is from trusted storage (not untrusted user input) - Original size is stored in header and validated - Pages have checksums for corruption detection Changes: - Added LZ4_DECOMPRESS_FAST method handle in FFILz4Compressor - Added decompressSegmentFast() method for fast mode - Updated decompressScoped() to use fast mode when enabled - Enabled by default in test configuration for benchmarking
Benchmarks showed LZ4_decompress_fast is actually ~25% SLOWER than LZ4_decompress_safe in LZ4 1.9.x. This is likely because the deprecated function now includes additional internal safety checks. Results: - Safe mode: 81s - Fast mode: 106s Updated default to use safe mode. Fast mode kept for testing only.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
Upgrades build/runtime to JDK 25 and adds extensive design/diagnostic documentation; no production source changes.
setup-javaand Docker base images to Temurin/Gradle JDK 25ai-docs/and planning files.gitignorefor build outputs, logs, and profiling artifactsWritten by Cursor Bugbot for commit 2b71dd1. This will update automatically on new commits. Configure here.