echo v0.1.1

[0.1.1] - 2026-05-26

• TrajectoryAccumulator better supports single and buffered timescales modes:
  • Buffered if: all leaves have a leading dim > 1
  • Single if: all leaves have leading dims == 1 or there is a single scalar
• Guide and API reference cover both accumulator modes and the detection rule.
• New project logo and README polish (#3).

Pull Requests

• docs: add logo by @sash-a in #3
• feat: better accumulator by @sash-a in #4

Full Changelog: v0.1.0...v0.1.1

echo v0.1.0

First public release.

echo exists because data movement was the bottleneck in our distributed RL pipelines. Existing options are good, but most are tied to a specific framework (e.g Reverb to TF) and allocate per-sample on the hot path. echo depends only on numpy and optree, and is built around a pre-allocated lockfree ring buffer in Rust that hands batches to Python as numpy views into Rust-owned memory. Designed for hundreds-to-thousands of concurrent rollout workers feeding a single learner.

In this release:

• Lockfree, pre-allocated ring buffer
• TCP transport with per-connection SPSC queues and a drainer pool
• Pytree samples (nested dicts/tuples of numpy arrays) via optree
• Zero-copy batches — sample() returns views, no copy after ingress
• GIL released while waiting on the next batch
• Per-batch backpressure and throughput metrics on SampleInfo
• Optional detailed-metrics build for histogram-based timing
• TrajectoryAccumulator helper for stacking rollout steps

Known limitations:

• FIFO is the only sampling strategy. Prioritised and uniform-random are on the list.
• TCP is the only transport for now; the interface is pluggable.
• Linux and macOS wheels only.

pip install id-echo

Docs: https://instadeepai.github.io/echo/