Skip to content

log-from-python-or-curl.md

Latest commit

 

History

History
149 lines (114 loc) · 4.54 KB

log-from-python-or-curl.md

File metadata and controls

149 lines (114 loc) · 4.54 KB
Raw

Log from Python (or curl, Go, anywhere)

Track runs from outside Clojure using chachaml's HTTP write API.

Goal

A non-Clojure team member should be able to log a run, log params/metrics, and end the run via REST. Useful for: Python training scripts, Go services, GitHub Actions, anywhere a quick curl is easier than a Clojure REPL.

Prerequisites

  • A running chachaml UI server (clojure -M:ui or the Docker setup). The write API lives under /api/w/.
  • Network access to the UI server from the client.

Steps

1. Bare curl, end-to-end

HOST=http://localhost:8080

# Start a run — returns the run id
RUN_ID=$(curl -s -X POST $HOST/api/w/runs \
  -H 'Content-Type: application/json' \
  -d '{"experiment":"iris","name":"from-python","tags":{"author":"maria"}}' \
  | jq -r .id)

echo "Run id: $RUN_ID"

# Log params
curl -X POST $HOST/api/w/runs/$RUN_ID/params \
  -H 'Content-Type: application/json' \
  -d '{"lr":0.01,"epochs":100}'

# Log a metric (single)
curl -X POST $HOST/api/w/runs/$RUN_ID/metrics \
  -H 'Content-Type: application/json' \
  -d '{"accuracy":0.94}'

# Log time-series metrics
for i in 0 1 2 3 4; do
  curl -X POST $HOST/api/w/runs/$RUN_ID/metrics \
    -H 'Content-Type: application/json' \
    -d "{\"loss\":$(echo "scale=3; 1.0 / ($i + 1)" | bc),\"step\":$i}"
done

# Optionally upload an artifact (raw bytes)
curl -X POST "$HOST/api/w/runs/$RUN_ID/artifacts?name=model.pkl" \
  -H 'Content-Type: application/octet-stream' \
  --data-binary @model.pkl

# End the run
curl -X POST $HOST/api/w/runs/$RUN_ID/end \
  -H 'Content-Type: application/json' \
  -d '{"status":"completed"}'

Verify in the UI: the run should appear in /runs with the params, metrics, and artifact.

2. From Python

import requests

HOST = "http://localhost:8080"
session = requests.Session()

# Start
r = session.post(f"{HOST}/api/w/runs",
                 json={"experiment": "iris",
                       "name": "from-python",
                       "tags": {"author": "maria"}})
r.raise_for_status()
run_id = r.json()["id"]

# Params
session.post(f"{HOST}/api/w/runs/{run_id}/params",
             json={"lr": 0.01, "epochs": 100})

# Metric per epoch
for epoch in range(100):
    session.post(f"{HOST}/api/w/runs/{run_id}/metrics",
                 json={"loss": 1.0 / (epoch + 1), "step": epoch})

# Final accuracy
session.post(f"{HOST}/api/w/runs/{run_id}/metrics",
             json={"accuracy": 0.94})

# End
session.post(f"{HOST}/api/w/runs/{run_id}/end",
             json={"status": "completed"})

3. (Optional) Register a model

curl -X POST $HOST/api/w/models \
  -H 'Content-Type: application/json' \
  -d "{\"name\":\"iris-classifier\",
       \"run_id\":\"$RUN_ID\",
       \"artifact\":\"model.pkl\",
       \"stage\":\"staging\",
       \"description\":\"first cut\"}"

The same exclusivity rule applies: promoting another version to :production demotes the previous one.

API surface

All endpoints accept and return JSON.

Verb + path Body Returns
POST /api/w/runs {experiment, name?, tags?, parent_run_id?} {id, ...}
POST /api/w/runs/:id/params {key: value, ...} {ok: true}
POST /api/w/runs/:id/metrics {key: value, step?: n} or {key1: v1, key2: v2} {ok: true}
POST /api/w/runs/:id/artifacts?name=... raw bytes {ok, sha256, size}
POST /api/w/runs/:id/end {status: "completed"|"failed", error?: "..."} {ok: true}
POST /api/w/models {name, run_id, artifact, stage?, description?} {name, version, ...}

Read endpoints are at /api/... (no w/ prefix); see Web UI tour: JSON API.

Troubleshooting

  • { "error": "no such run" } from /params or /metrics — you're posting to a run id that doesn't exist or was already ended. Track the run id from the /runs response and stop logging after /end.
  • Content-Type mismatches — params/metrics endpoints expect JSON; artifact uploads expect application/octet-stream. Mixing them returns a 400.
  • Performance — each metric POST is one HTTP round-trip. For high-frequency time-series (e.g. logging per training step in a tight loop), batch metrics into one call: {"loss": 0.4, "accuracy": 0.9, "step": 12} logs both at the same step.

Where to go next

  • For higher-throughput pipelines, prefer the Clojure API directly — it skips the HTTP layer entirely.
  • For language bindings, the API is small enough that any HTTP client works. There's no official Python package, by design — the REST surface is the contract.