Performance

The service avoids requiring each integration to install and start conda. Its latency depends on metadata freshness, worker state and whether a complete result is retained.

Work performed

A miss can include input parsing, repodata download or validation, index construction, SAT solving, export and serialization. A one-shot CLI also pays Python and conda startup costs. A persistent worker retains indexes from Presto’s ordinary solve path between requests. Workspace solves call conda-workspaces and conda’s solver API, so they do not use that direct index cache. Without persistent-worker mode, HTTP uses an isolated process for each uncached solve.

A full result hit skips solving and rendering after validating the request’s metadata state. /r/<hash> is cheaper still because it retrieves stored bytes without that freshness check. Measure these operations separately.

Concurrency

Ordinary solves use separate processes for multiple target platforms. More workers can reduce elapsed time while increasing memory use. Workspace solves process selected environment/target pairs sequentially in the solve process. CONDA_PRESTO_WORKERS does not parallelize those pairs.

The default container serializes foreground requests through one persistent worker, whose internal pool can still solve multiple ordinary platforms. Workspace target count and ordinary platform count therefore have different effects on latency.

Additional service instances provide independent process state. Shared Redis makes retained outputs available to those instances, but local metadata markers can give equivalent requests different cache keys. One-instance and multiple-instance throughput must be observed on the same workload and channel content. There is no published service-level performance guarantee.

What to record

Report explicit specs, channels, platforms, component versions, virtual-package overrides, metadata freshness, worker reuse, cache backend, concurrency and sample count. For workspaces, also report selected environments and logical targets, their subdirs and manifest system requirements. Compare cold startup, misses and retained results separately. Faster rattler solving reduces miss cost, which can narrow the advantage of result reuse.

Supported HTTP lock conversions and CLI --export conversions perform no SAT solving or package download and should be measured separately from solves.

See Benchmark the service for controlled measurements and Cache reference for cache behavior.