v0.2.3

Key Changes

• New approaches: feddyn, oort, FedBalancer, synchronous hierarchical FL with coordinator, coordinated asyncfl, and differential privacy
• Metrics collection support: collect cpu/gpu utilization, ram and vram usage, bytes sent and received and the time to execute functions
• Local registry support: model and metrics are saved in a local directory, which allows experiments without mlflow
• misc improvements:
  • support for channel leave functionality in the p2p backend
  • support for simplifying tasklet composition primitives
  • protobuf package version upgrade to be compatible with tensorflow v2.12.0
  • deployer: fix bug to update pod status incorrectly
  • fix a bug that removes model updates from a disk cache when the total cache data is larger than its size
  • fix openapi specification mismatch between a released image and the source

What's Changed

• example/implementation for Oort by @jaemin-shin in #369
• feddyn implementation pytorch by @GustavBaumgart in #370
• update for Oort: overcommitment support by @jaemin-shin in #372
• fix: remove non-orchestration mode by @openwithcode in #374
• update for Oort: accurate calculation of round duration with timestamp by @jaemin-shin in #375
• feat: replica support for roles by @openwithcode in #376
• fix: remove errors regarding group association to SDK config files by @jaemin-shin in #378
• doc: update supported algorithms/mechanisms by @myungjin in #377
• fix: missing delta weights computation in middle aggregator by @myungjin in #382
• fix: incorrect groupAssociation in lib example configs by @myungjin in #383
• refactor: move syncfl files by @myungjin in #384
• Synchronous orchestrator architecture by @elqurio in #379
• feat: per-channel backend support by @myungjin in #381
• fix: hardening coordinated syncfl by @myungjin in #385
• refactor: auto-formatting sdk files with black by @myungjin in #386
• misc: example for synchronous hierachical FL with coordinator by @myungjin in #388
• feat+refactor+fix: per-channel backend support in SDK by @myungjin in #387
• fix: zero weight upload from trainer in hybrid mode by @jaemin-shin in #389
• feat: optimize composer's extensibility by @myungjin in #390
• Remove unnecessary config examples by @elqurio in #391
• example/implementation for FedBalancer, with a new sampler category by @jaemin-shin in #380
• refactor: remove legacy code by @myungjin in #393
• feat: coordination revision by @myungjin in #394
• updated feddyn implementation pytorch by @GustavBaumgart in #392
• chore: syncfl_hier_coord_mnist schema update by @myungjin in #395
• feat: channel leave functionality for p2p backend by @myungjin in #398
• feat: additional optimization for channel leave by @myungjin in #399
• Impove tasklet composition extensibility by @elqurio in #397
• Move examples to the root of the lib/python dir by @elqurio in #400
• feddyn algorithm correction by @GustavBaumgart in #401
• fix: slow tx task termination by @myungjin in #402
• Fix alias already exists for coord_syncfl by @elqurio in #406
• Monitor deployed tasks and update status in case of crashes by @openwithcode in #404
• new feature: add differential privacy by @jaemin-shin in #403
• feat: coordinated asyncfl by @myungjin in #407
• fix: Resolve the get tasks command by job ID to return the expected tasks details of a job by @openwithcode in #412
• fix: refactor examples tests by @openwithcode in #409
• bump: upgrade version of go to 1.18 by @openwithcode in #411
• feat: track tasklet runtime by @GustavBaumgart in #405
• feat: communication cost metric collection by @GustavBaumgart in #414
• documentation: fix typos in README.md and docs by @jaemin-shin in #416
• doc: add citation for flame arxiv paper by @myungjin in #417
• refactor: optimize create job task of control plane by @jaemin-shin in #415
• fix: disable automatic eviction of diskcache at aggregation by @jaemin-shin in #419
• feat: utilization/memory usage on CPU and GPU by @GustavBaumgart in #420
• fix: GPU/CPU monitoring termination by @GustavBaumgart in #424
• fix: protobuf version error by @GustavBaumgart in #425
• feat: local registry by @GustavBaumgart in #421
• fix: fix oort utility calculation and add medmnist example for oort by @jaemin-shin in #423
• refactor: monitoring scope adjustment in deployer by @myungjin in #427
• feat: local registry timestamp by @GustavBaumgart in #426
• misc: add debug message in deployer by @myungjin in #428
• fix: incorrect task status update by deployer by @myungjin in #429
• fix: check payload before accumulating bytes received by @GustavBaumgart in #430
• fix: deployer pod state update by @myungjin in #431

New Contributors

• @jaemin-shin made their first contribution in #369

Full Changelog: v0.2.2...v0.2.3