v0.2.0a2

This is the release notes of v0.2.0a2. See here for the complete list of solved issues and merged PRs.

Highlights

• Eager mode support, if eager mode is on, tensor will be executed immediately once it's created.
• Initial worker-level failover support.

New Features

• Replace executed tensor with fetch tensor (#180)
• Support eager mode (#200)
• Support sparse for some operators in mt.linalg (#215)
• Add TileDB support as external storage (#227)
• Add worker-level fail-over support for single graph (#207)

Enhancements

• Pre-push operands into worker when the worker can be estimated (#119)
• Separate graph analyzer from graph and add test cases (#193)
• Fix codacy problems (#204)
• Refactor executor to do operand-oriented execution and support actor operation (#212)
• Add __mars_tokenize__ protocol (#218)
• Manually add methods for Tensor class (#223)
• Use random object id in plasma store to avoid potential collision (#226)
• Assign initials given existing data in eager mode (#242)

Bug fixes

• Fix serializing graph with a tiled tensor (#191)
• Fix starting with multiple schedulers with --schedulers (#202)
• Fix process clean up when local cluster terminates (#206)
• Fix bug that TensorZeros generate same key even if their shapes are different in TensorDiag.tile (#210)
• Fix bug that chunks of ones and zeros with same shape generated different key (#216)
• Fix execute one of the multiple output tensors (#221)
• Add TensorInv operand to fix bug in mt.linalg.inv (#237)
• Fix executor's behavior of result deletions (#249)

Documentation

• Add contributing guide & remove tracing on tags (#181)

Installation

• Use fixed versions of dependencies to build wheels (#198)

Tests

• Fix CI failure under python 2.7 in MacOS (#188)
• Use conda for all configurations (#229)