Benchmark performance#

Use a controlled profile to determine which part of a conda-presto request is expensive for your workload. Do not combine cold startup, refreshed repodata, index reuse, and full result-cache hits into one result.

Important

Run comparisons on the same host against the same channels and repodata state. A result from another day can represent different package metadata and a different solve.

Run the repository benchmarks#

From a source checkout, install the test environment and run the benchmark task:

pixi install --environment test
pixi run -e test bench

The suite measures serialization paths, representative single-platform solves, and a multi-platform solve. Solve benchmarks can access channel metadata, so record whether repodata was downloaded or refreshed during the run.

Select an HTTP profile#

Run one profile at a time in an activated installation or a source-checkout Pixi shell. Restart the server after changing a profile.

Full result-cache hit

Keep the result cache enabled. Ordinary HTTP mode is sufficient because a hit skips the isolated solve process.

export CONDA_PRESTO_RESULT_CACHE_BACKEND=memory
export CONDA_PRESTO_RESULT_CACHE_SIZE=256
export CONDA_PRESTO_PERSISTENT_WORKER=0

Ordinary HTTP miss

Disable result retention to measure the isolated child process used by each uncached ordinary HTTP request.

export CONDA_PRESTO_RESULT_CACHE_BACKEND=memory
export CONDA_PRESTO_RESULT_CACHE_SIZE=0
export CONDA_PRESTO_PERSISTENT_WORKER=0

Persistent-worker miss

Disable result retention while keeping the worker alive. Repeated requests can reuse its loaded rattler indexes but still run the solve.

export CONDA_PRESTO_RESULT_CACHE_BACKEND=memory
export CONDA_PRESTO_RESULT_CACHE_SIZE=0
export CONDA_PRESTO_PERSISTENT_WORKER=1

Keep the requested startup set narrow, disable rate limiting for the local measurement, and start the server:

export CONDA_PRESTO_CHANNELS=conda-forge
export CONDA_PRESTO_ALLOWED_CHANNELS=conda-forge
export CONDA_PRESTO_PLATFORMS=linux-64
export CONDA_PRESTO_RATE_LIMIT=0

conda presto --serve --host 127.0.0.1 --port 8000

Wait for readiness in another terminal:

until curl --fail --silent http://127.0.0.1:8000/health > /dev/null
do
  sleep 1
done

Prime the selected state#

Send one request before collecting samples:

curl --fail --silent --show-error \
  --get http://127.0.0.1:8000/resolve \
  --data-urlencode 'spec=python=3.13' \
  --data-urlencode 'spec=numpy' \
  --data-urlencode 'channel=conda-forge' \
  --data-urlencode 'platform=linux-64' \
  --output /dev/null

For the full result-cache profile, this creates the entry that later requests can reuse. For the two miss profiles, it stabilizes the on-disk repodata state. Persistent-worker startup has already built the configured worker index.

Collect repeated HTTP timings#

Record loopback request time without writing response bodies to disk:

for run in 1 2 3 4 5 6 7 8 9 10
do
  curl --fail --silent --show-error \
    --get http://127.0.0.1:8000/resolve \
    --data-urlencode 'spec=python=3.13' \
    --data-urlencode 'spec=numpy' \
    --data-urlencode 'channel=conda-forge' \
    --data-urlencode 'platform=linux-64' \
    --output /dev/null \
    --write-out "$run %{time_total}\n"
done | tee timings.txt

Calculate the median and range with the standard library:

python - <<'PY'
from pathlib import Path
from statistics import median

samples = [
    float(line.split()[1])
    for line in Path("timings.txt").read_text().splitlines()
]
print(f"samples: {len(samples)}")
print(f"median: {median(samples):.6f} s")
print(f"range: {min(samples):.6f} to {max(samples):.6f} s")
PY

Repeat the procedure for each profile. Treat GET /r/<hash> as a separate operation. It retrieves a stored snapshot without the repodata freshness check performed by a repeated /resolve request.

Report enough context#

Include these details with the result:

execution mode and process warmth
package specs, channels, and platforms
conda, conda-presto, and solver versions
repodata TTL and whether a refresh occurred
cache backend and the intended hit or miss profile
prefix state and virtual-package overrides for solver-backend tests
median, range, and number of samples

Use Performance to interpret the cost model and Cache reference to distinguish result-cache reuse from index reuse.