aboutsummaryrefslogtreecommitdiff
path: root/src/zenstore/compactcas.cpp
Commit message (Collapse)AuthorAgeFilesLines
* long filename support (#330)Dan Engelbrecht2025-03-311-15/+15
| | | - Bugfix: Long file paths now works correctly on Windows
* zen build cache service (#318)Dan Engelbrecht2025-03-261-2/+2
| | | | | | | | | - **EXPERIMENTAL** `zen builds` - Feature: `--zen-cache-host` option for `upload` and `download` operations to use a zenserver host `/builds` endpoint for storing build blob and blob metadata - Feature: New `/builds` endpoint for caching build blobs and blob metadata - `/builds/{namespace}/{bucket}/{buildid}/blobs/{hash}` `GET` and `PUT` method for storing and fetching blobs - `/builds/{namespace}/{bucket}/{buildid}/blobs/putBlobMetadata` `POST` method for storing metadata about blobs - `/builds/{namespace}/{bucket}/{buildid}/blobs/getBlobMetadata` `POST` method for fetching metadata about blobs - `/builds/{namespace}/{bucket}/{buildid}/blobs/exists` `POST` method for checking existance of blobs
* Miscellaneous minor LLM fixes (#268)v5.5.17-pre0Stefan Boberg2024-12-171-1/+2
| | | | | | | With this change, LLM tags are assigned using the name,parent tuple rather than just by name only. This allows tag hierarchies like `cache/store` and `project/store` which would previously get collapsed into the first pair seen when registering the `store` tag. This PR also adds some more LLM tag annotations to more accurately associate memory allocations with subsystems In addition, this PR also tweaks the frequency of timer marker events to increase the resolution in Insights and avoid some cases of Insights deciding that marker events are too far apart since we don't allocate as frequently as UE tends to.
* batch fetch record cache values (#266)Dan Engelbrecht2024-12-171-17/+18
| | | | | | - Improvement: Batch fetch record attachments when appropriate - Improvement: Reduce memory buffer allocation in BlockStore::IterateBlock - Improvement: Tweaked BlockStore::IterateBlock logic when to use threaded work (at least 4 chunks requested) - Bugfix: CasContainerStrategy::IterateChunks could give wrong payload/index when requesting 1 or 2 chunks
* projectstore getchunks rpc with modtag (#244)Dan Engelbrecht2024-12-051-15/+17
| | | Feature: Project store "getchunks" rpc call /prj/{project}/oplog/{log}/rpc extended to accept both CAS (RawHash) and Id (Oid) identifiers as well as partial ranges
* Unity build fixes (#253)Stefan Boberg2024-12-051-13/+13
| | | some fixes to make everything build using unity build mode. Mostly moved code from anonymous namespaces into local impl namespace to avoid ambiguity in name resolution.
* added support for dynamic LLM tags (#245)Stefan Boberg2024-12-021-0/+34
| | | | | * added FLLMTag which can be used to register memory tags outside of core * changed `UE_MEMSCOPE` -> `ZEN_MEMSCOPE` for consistency * instrumented some subsystems with dynamic tags
* use plain sorted array instead of map of vectors (#237)Dan Engelbrecht2024-11-271-4/+7
| | | | | * use plain sorted array instead of map of vectors * reserve vectors up front = 5% perf increase * don't do batch read of chunks if we have a single chunk -> 1% perf gain
* caller controls threshold for bulk-loading chunks in IterateChunks (#222)Dan Engelbrecht2024-11-251-3/+5
| | | | | | * Allow caller to control threshold for bulk-loading chunks in IterateChunks * use smaller batch chunk reading for /fileinfos and /chunkinfos as we do not intend to read the payload * use smaller batch read buffer when just querying for size of attachments
* stronger validation of payload existance (#229)Dan Engelbrecht2024-11-251-2/+26
| | | | | | - Don't add RawSize and Size in ProjectStore::GetProjectFiles response if we can't get the payload - Use validation of payload size/existance in all chunk fetch operations in file cas - In project store oplog validate, make sure we can reach all the payloads - Add threading to oplog validate request
* fix inconsistencies in filecas due to failing to remove payload file during ↵Dan Engelbrecht2024-11-221-1/+1
| | | | | | | | GC (#224) make sure we rewrite filecas entries if chunk size changes (due to compression changes) hardening of move/write files in filecas if we encounter a filecas entry with mismatching size (due to pre-existing bug) we validate the file and update the index if we find a bad filecas file on disk we now attempt to remove it
* remove gc v1 (#121)Dan Engelbrecht2024-10-031-610/+29
| | | | | * kill gc v1 * block use of gc v1 from zen command line * warn and flip to gcv2 if --gc-v2=false is specified for zenserver
* gc block size target max size (#180)Dan Engelbrecht2024-10-021-28/+25
| | | | | | * If a block is small (less than half max size) we add it to blocks to compact Sort blocks when iterating over them * do compact of block stores even if no new unused are found * do compact phase even if bucket is empty
* optimize startup time (#175)Dan Engelbrecht2024-09-301-3/+3
| | | | | | * use tsl::robin_set for BlockIndexSet don't calculate full block location when only block index is needed * don't copy visitor function * reserve space for attachments
* Add `gc-attachment-passes` option to zenserver (#167)Dan Engelbrecht2024-09-251-1/+1
| | | | | Added option `gc-attachment-passes` to zenserver Cleaned up GCv2 start and stop logs and added identifier to easily find matching start and end of a GC pass in log file Fixed project store not properly sorting references found during lock phase
* gc unused refactor (#165)Dan Engelbrecht2024-09-231-5/+9
| | | | | * optimize IoHash and OId comparisions * refactor filtering of unused references * add attachment filtering to gc
* gc performance improvements (#160)Dan Engelbrecht2024-09-171-13/+14
| | | | | | | | | | * optimized ValidateCbUInt * optimized iohash comparision * replace unordered set/map with tsl/robin set/map in blockstore * increase max buffer size when writing cache bucket sidecar * only store meta data for files < 4Gb * faster ReadAttachmentsFromMetaData * remove memcpy call in BlockStoreDiskLocation * only write cache bucket state to disk if GC deleted anything
* trace scopes improvementsDan Engelbrecht2024-09-101-0/+1
|
* move gc logs to gc logger (#142)Dan Engelbrecht2024-09-041-0/+6
| | | - Improvement: Move GC logging in callback functions into "gc" context
* stop exceptions from leaking on threaded work (#102)Dan Engelbrecht2024-08-061-23/+33
| | | | * catch exceptions in threaded work * don't abort all project file/chunk info fetch for single failure
* refactor BlockStore IterateChunks (#77)Dan Engelbrecht2024-05-171-18/+55
| | | Improvement: Refactored IterateChunks to allow reuse in diskcachelayer and hide public GetBlockFile() function in BlockStore
* use direct file access for large file hash (#63)Dan Engelbrecht2024-04-261-4/+4
| | | - Improvement: Refactor `IoHash::HashBuffer` and `BLAKE3::HashBuffer` to not use memory mapped files. Performs better and saves ~10% of oplog export time on CI
* iterate cas chunks (#59)Dan Engelbrecht2024-04-241-10/+43
| | | - Improvement: Reworked GetChunkInfos in oplog store to reduce disk thrashing and improve performance
* safer gcv2 on error (#60)Dan Engelbrecht2024-04-241-0/+2
| | | - Bugfix: Harden GCv2 when errors occur and gracefully abort GC operation on error
* InsertChunks for CAS store (#55)Dan Engelbrecht2024-04-221-0/+58
| | | - Improvement: Add batching when writing multiple small chunks to block store - decreases I/O load significantly on oplog import
* improved assert (#37)Dan Engelbrecht2024-04-041-2/+2
| | | | - Improvement: Add file and line to ASSERT exceptions - Improvement: Catch call stack when throwing assert exceptions and log/output call stack at important places to provide more context to caller
* fix potential partially written files (#2)Dan Engelbrecht2024-03-131-3/+12
| | | | * Make sure WriteFile() does not leave incomplete files * use TemporaryFile and MoveTemporaryIntoPlace to avoid leaving partial files on error
* Make sure we wait for all scheduled tasks to complete before throwing ↵Dan Engelbrecht2024-02-281-2/+6
| | | | | exceptions further (#662) Bugfix: We must not throw exceptions to calling function until all async work we spawned has returned
* improved block store logging and more gcv2 tests (#659)Dan Engelbrecht2024-02-271-50/+96
| | | | * improved gc/blockstore logging * more gcv2 tests
* fix ChunkIndexToChunkHash indexing (#621)Stefan Boberg2023-12-191-1/+1
| | | would previously index into a reserved-but-not-sized vector which is bad but not crash-inducing bad
* improve trace (#606)Dan Engelbrecht2023-12-131-0/+1
| | | | | * Adding some more trace scopes for better visiblity * Removed spammy trace scope when replaying oplogs * Remove "::Disk" from trace scopes - redundant now that we have merge disk and memory layers
* improved scrubbing of oplogs and filecas (#596)Stefan Boberg2023-12-111-26/+12
| | | | | | - Improvement: Scrub command now validates compressed buffer hashes in filecas storage (used for large chunks) - Improvement: Added --dry, --no-gc and --no-cas options to zen scrub command - Improvement: Implemented oplog scrubbing (previously was a no-op) - Improvement: Implemented support for running scrubbint at startup with --scrub=<options>
* fixed bug in CasContainerStrategy::ReadIndexFile (#595)Stefan Boberg2023-12-071-1/+3
| | | this was introduced in a recent optimization and would cause CAS items to not be found after a shutdown/restart cycle
* reserve vectors in gcv2 upfront / load factor for robin_map (#582)Dan Engelbrecht2023-12-041-6/+15
| | | | | * reserve vectors in gcv2 upfront * set max load factor for robin_map indexes to reduce memory usage * set min load factor for robin_map indexes to allow them to shrink
* use 32 bit offset and size in BlockStoreLocation (#581)Dan Engelbrecht2023-12-011-1/+1
| | | - Improvement: Reduce memory usage in GC and diskbucket flush
* tracing for gcv2 (#574)Dan Engelbrecht2023-11-281-0/+6
| | | | | | - Improvement: Added more trace scopes for GCv2 - Bugfix: Make sure we can override flags to "false" when running `zen gc` commmand - `smallobjects`, `skipcid`, `skipdelete`, `verbose`
* fix missing locks/sync of log position when writing index snapshots (#572)Dan Engelbrecht2023-11-271-2/+4
| | | | * fix missing locks/sync of log position when writing index snapshots * changelog
* optimized index snapshot reading/writing (#561)Stefan Boberg2023-11-271-18/+35
| | | | | the previous implementation of in-memory index snapshots serialise data to memory before writing to disk and vice versa when reading. This leads to some memory spikes which end up pushing useful data out of system cache and also cause stalls on I/O operations. this change moves more code to a streaming serialisation approach which scales better from a memory usage perspective and also performs much better
* Add GC Cancel/Stop (#568)Dan Engelbrecht2023-11-241-0/+19
| | | | - GcScheduler will now cancel any running GC when it shuts down. - Old GC is rather limited in *when* it reacts to cancel of GC. GCv2 is more responsive.
* reduce work when there are no blocks to compact (#558)Dan Engelbrecht2023-11-221-51/+54
| | | | * reduce work when there are no blocks to compact * fix lock scopes
* add command line options for compact block threshold and gc verbose (#557)Dan Engelbrecht2023-11-211-7/+13
| | | | | | | | | | | - Feature: Added new options to zenserver for GC V2 - `--gc-compactblock-threshold` GCV2 - how much of a compact block should be used to skip compacting the block, default is 90% - `--gc-verbose` GCV2 - enable more verbose output when running a GC pass - Feature: Added new options to `zen gc` command for GC V2 - `--compactblockthreshold` GCV2 - how much of a compact block should be used to skip compacting the block, default is 90% - `--verbose` GCV2 - enable more verbose output when running a GC pass - Feature: Added new parameters for endpoint `admin/gc` (PUT) - `compactblockthreshold` GCV2 - how much of a compact block should be used to skip compacting the block, default is 90% - `verbose` GCV2 - enable more verbose output when running a GC pass
* compact separate for gc referencer (#533)Dan Engelbrecht2023-11-211-113/+106
| | | | | - Refactor GCV2 so GcReferencer::RemoveExpiredData returns a store compactor, moving out the actual disk work from deleting items in the index. - Refactor GCV2 GcResult to reuse GcCompactStoreStats and GcStats - Make Compacting of stores non-parallell to not eat all the disk I/O when running GC
* disk layer gc and error/warnings cleanup (#515)Dan Engelbrecht2023-11-081-2/+2
| | | | | | | - Improvement: Use GC reserve when writing index/manifest for a disk cache bucket when disk is low when available - Improvement: Demote errors to warning for issues that are not critical and we handle gracefully - Improvement: Treat more out of memory errors from windows as Out Of Memory errors Fixed wrong sizeof() statement for compactcas index (luckily the two structs are of same size)
* gc v2 tests (#512)Dan Engelbrecht2023-11-061-27/+9
| | | | | | | | | | * set MaxBlockCount at init * properly calculate total size * basic blockstore compact blocks test * correct detection of block swap * Use one implementation for CreateRandomBlob * reduce some data sets to increase speed of tests * reduce test time * rename BlockStoreCompactState::AddBlock -> BlockStoreCompactState::IncludeBlock
* individual gc stats (#506)Dan Engelbrecht2023-10-301-33/+56
| | | | | - Feature: New parameter for endpoint `admin/gc` (GET) `details=true` which gives details stats on GC operation when using GC V2 - Feature: New options for zen command `gc-status` - `--details` that enables the detailed output from the last GC operation when using GC V2
* New GC implementation (#459)Dan Engelbrecht2023-10-301-0/+217
| | | - Feature: New garbage collection implementation, still in evaluation mode. Enabled by `--gc-v2` command line option
* Remove any unreferenced blocks in block store on open (#492)Dan Engelbrecht2023-10-231-1/+1
| | | * Remove any unreferenced blocks in block store on open
* Don't prune block locations due to missing blocks a startup (#487)Dan Engelbrecht2023-10-201-46/+5
| | | | | | * Don't prune block locations due to missing blocks a startup This makes the behaviour consistent with FileCas - you can have an index that is not fully backed by data. Asking for a location that is not backed by data results in getting an empty result back Also, don't try to GC blocks that are unknown to the block store at the time of snapshot (to avoid removing data that comes in after GatherReferences in GC)
* clean up GcContributor and GcStorage to be pure interfaces (#485)Dan Engelbrecht2023-10-201-1/+3
|
* Add --skip-delete option to gc command (#484)Dan Engelbrecht2023-10-201-1/+1
| | | | - Feature: Add `--skip-delete` option to gc command - Bugfix: Fix implementation when claiming GC reserve during GC
generated by cgit v1.2.3 (git 2.25.1) at 2026-03-12 05:26:19 +0000