Jeremie Fraeys
a5c1a9fc0b
refactor: Phase 4 - split worker package into focused files
...
Split 551-line worker/core.go into single-concern files:
- worker/config.go (+44 lines)
- Added config parsing: envInt(), parseCPUFromConfig(), parseGPUCountFromConfig()
- parseGPUSlotsPerGPUFromConfig()
- Now has all config logic in one place (440 lines total)
- worker/metrics.go (new file, 172 lines)
- Extracted setupMetricsExporter() with ~30 Prometheus metric registrations
- Isolated metrics logic for easy modification
- worker/factory.go (new file, 183 lines)
- Extracted NewWorker() factory function
- Moved prePullImages(), pullImage() from core.go
- Centralized worker instantiation
- worker/worker.go (renamed from core.go, ~100 lines)
- Now just defines Worker struct, MLServer, JupyterManager
- Clean, focused file without mixed concerns
Lines redistributed: ~350 lines moved from monolithic core.go
Build status: Compiles successfully
2026-02-17 12:57:02 -05:00
Jeremie Fraeys
d1bef0a450
refactor: Phase 3 - fix config/storage boundaries
...
Move schema ownership to infrastructure layer:
- Redis keys: config/constants.go -> queue/keys.go (TaskQueueKey, TaskPrefix, etc.)
- Filesystem paths: config/paths.go -> storage/paths.go (JobPaths)
- Create config/shared.go with RedisConfig, SSHConfig
- Update all imports: worker/, api/helpers, api/ws_jobs, api/ws_validate
- Clean up: remove duplicates from queue/task.go, queue/queue.go, config/paths.go
Build status: Compiles successfully
2026-02-17 12:49:53 -05:00
Jeremie Fraeys
a93b6715fd
feat: add native library bridge and queue integration
...
- Add native_queue.go with CGO bindings for queue operations
- Add native_queue_stub.go for non-CGO builds
- Add hash_selector to choose between Go and native implementations
- Add native_bridge_libs.go for CGO builds with native_libs tag
- Add native_bridge_nocgo.go stub for non-CGO builds
- Update queue errors and task handling for native integration
- Update worker config and runloop for native library support
2026-02-16 20:38:30 -05:00
Jeremie Fraeys
43d241c28d
feat: implement C++ native libraries for performance-critical operations
...
- Add arena allocator for zero-allocation hot paths
- Add thread pool for parallel operations
- Add mmap utilities for memory-mapped I/O
- Implement queue_index with heap-based priority queue
- Implement dataset_hash with SIMD support (SHA-NI, ARMv8)
- Add runtime SIMD detection for cross-platform correctness
- Add comprehensive tests and benchmarks
2026-02-16 20:38:04 -05:00
Jeremie Fraeys
d408a60eb1
ci: push all workflow updates
Documentation / build-and-publish (push) Waiting to run
Test / test (push) Waiting to run
Checkout test / test (push) Successful in 5s
CI with Native Libraries / test-native (push) Has been cancelled
CI with Native Libraries / build-release (push) Has been cancelled
2026-02-12 13:28:15 -05:00
Jeremie Fraeys
2e701340e5
feat(core): API, worker, queue, and manifest improvements
...
- Add protocol buffer optimizations (internal/api/protocol.go)
- Add filesystem queue backend (internal/queue/filesystem_queue.go)
- Add run manifest support (internal/manifest/run_manifest.go)
- Worker and jupyter task refinements
- Exported test wrappers for benchmarking
2026-02-12 12:05:17 -05:00
Jeremie Fraeys
72b4b29ecd
perf: add profiling benchmarks and parallel Go baseline for C++ optimization
...
Add comprehensive benchmarking suite for C++ optimization targets:
- tests/benchmarks/dataset_hash_bench_test.go - dirOverallSHA256Hex profiling
- tests/benchmarks/queue_bench_test.go - filesystem queue profiling
- tests/benchmarks/artifact_and_snapshot_bench_test.go - scanArtifacts/extractTarGz profiling
- tests/unit/worker/artifacts_test.go - moved from internal/ for clean separation
Add parallel Go implementation as baseline for C++ comparison:
- internal/worker/data_integrity.go: dirOverallSHA256HexParallel() with worker pool
- Benchmarks show 2.1x speedup (3.97ms -> 1.90ms) vs sequential
Exported wrappers for testing:
- ScanArtifacts() - artifact scanning
- ExtractTarGz() - tar.gz extraction
- DirOverallSHA256HexParallel() - parallel hashing
Profiling results (Apple M2 Ultra):
- dirOverallSHA256Hex: 78% syscall overhead (target for mmap C++)
- rebuildIndex: 96% syscall overhead (target for binary index C++)
- scanArtifacts: 87% syscall overhead (target for fast traversal C++)
- extractTarGz: 95% syscall overhead (target for parallel gzip C++)
Related: C++ optimization strategy in memory 5d5f0bb6
2026-02-12 12:04:02 -05:00
Jeremie Fraeys
82034c68f3
feat(worker): add integrity checks, snapshot staging, and prewarm support
2026-01-05 12:31:13 -05:00
