v0.17-beta

Changelog:

• remove the cross-service transaction protocol's requirement for bidirectional synchronization messages
• break the startup process into multiple stages
• allow partial replication without joining a cluster
• implement support for local peer discovery via mDNS
• upgrade to hyper 1.4
• require a correct cluster state hash to join a cluster
• gate Chain and each Collection type with a feature flag for fast debug builds
• separate the tc_server sub-crate from the main crate

v0.16.0 beta

Changelog:

• validate auth tokens exactly once, upon receipt
• reject unauthorized requests without reading payload data (headers only)
• remove unnecessary lock acquisitions when instantiating a new request handler
• move per-transaction garbage collection completely into a background task
• separate a new tinychain_async package out of the tinychain Python client
• move client implementations into a single client directory

v0.15.0 beta

Changelog:

• improved OpenCL support & performance

This release re-introduces a separate binary which supports GPU acceleration via OpenCL.

v0.14.0 beta

Changelog:

• improved table table index merge performance
• improved memory performance by minimizing heap allocations

0.13.1 beta

This release includes incremental bug fixes and performance improvements focused on supporting high-traffic applications.

0.13.0-beta

Version 0.13-beta introduces experimental support for multi-GPU hardware acceleration using the ha-ndarray library. If you are using TinyChain, we want to hear from you! Send an email with the word "hello" to [email protected].

Changelog:

• dropped ArrayFire as a dependency
• added experimental support for multi-GPU hardware acceleration (build with the "opencl" feature flag to enable this)
• updated collection types to implement a log-structured merge consistency model, for faster commits
• decoupled transactional locks from filesystem locks
• added support for cell level locking
• added support for multiple axes in tensor reduce operations (including einsum)
• updated crate structure to include the new chain, collection, fs, scalar, and state sub-crates for faster build times

0.12.0 beta

changelog:

• add support for installing a new Cluster at runtime without restarting
• implement basic support for /class, /library, and /service endpoints
• require authorization to install a new Cluster
• performance improvements to transactional locking
• replace App with Service in the Python client
• update linear algebra and ml services and tests to use Service with dynamic installation
• refactor Chain to support a generic mutable subject
• refactor tc_transact::File to support a generic key type
• bugfixes to Chain recovery in single-host mode
• security updates to dependency versions

SHA256 checksum: 9cf7d9408f3e10e8d1dbc92f6cd8ccf6bffd24992e82dfbe842b48c23266b87f

0.11.0 beta

This release focuses on host performance.

Changelog:

• chain performance improvements
• transactional filesystem performance improvements
• transaction lock performance improvements
• complex number performance improvements
• gradient support for for Tensor.max, Tensor.min, Tensor.product, Tensor.sum

SHA256 checksum: b15b4f7734946ff2c3303b84bc8f0bcdc3a8b213317e731215c6156d4583cc8e

0.10.0 beta

The 0.10.0 beta release focuses on client usability.

Changelog:

• dropped support for Python versions < 3.9
• added support for generic types (e.g. in Map, Tuple, Op, etc)
• added support for deriving Ops via reflection (i.e. reflecting over an existing user-defined Op to create a new, different Op)
• added support for differentiable Ops, including instance methods
• refactored NeuralNet and Layer classes to use differentiable methods
• Graph schema can now be auto-generated by providing a list of Models
• added explicit support for the multivariate chain rule, for automatic differentiation
• refactored Context to require only unique states and support constructing a new Context based on the structure of an existing Context
• added hash_of for consistent hashes of TinyChain States
• refactored URI to replace MethodSubject

SHA256 checksum: 6a019e90af5b5df4cb1b3d74c42f73d6135c3a7efb753b718137573c147e3098

0.9.0 beta

Changelog:

• added support for automatic differentiation of Tensor transforms (e.g. Tensor.slice, Tensor.reshape)
• improved TxnLock performance & reliability under high-load conditions
• added support for computing partial derivatives w/r/t any State in a differentiable operator graph
• improved support for referencing a differentiable operator graph from an Op context
• added support for custom gradient definitions
• implemented Tensor.max and Tensor.min
• improved einsum performance
• added a keepdims option to Tensor reduce methods
• improved compile-time tracking of tensor shapes
• refactored testutils for use in client application tests

SHA256 checksum: 7623181e63524389ecc0f82537556f699757be961feb9181f54c86d0d0a84e9c