datarobot-workload-api

DataRobot Workload API

Run container images as managed, autoscalable services on DataRobot. One skill, four jobs — pick the section by user intent:

1. Create / configure / scale — deploy a container; change replicas, resources, autoscaling, bundle; inject credentials
2. Diagnose — workload is stuck, errored, or crash-looping
3. Observe — logs, traces, metrics, service stats for a running workload
4. Artifact lifecycle — iterate drafts, build images, lock for production, roll out new versions

Prerequisites

DATAROBOT_ENDPOINT (must end in /api/v2) and DATAROBOT_API_TOKEN must be set. Run datarobot-setup if not. Auth header: Authorization: Bearer ${DATAROBOT_API_TOKEN}. The Workload API is not in the datarobot Python SDK — call REST directly.

Transport. Examples use Python httpx (pip install httpx). The API is plain HTTP, so equivalent calls work via curl or the pulumi-datarobot Pulumi provider declaratively. The skill teaches the model; transport is interchangeable.

Bundled scripts

Runnable Python in scripts/ (this skill's folder). Each uses httpx and reads DATAROBOT_ENDPOINT + DATAROBOT_API_TOKEN:

• wait_for_running.py <workload_id> — poll until running; exit 2 on terminal failure, 3 on timeout
• diagnose_workload.py <workload_id> — run the 5-step debug flow, print a structured diagnosis (--json for machine-readable)
• wait_for_build.py <artifact_id> <build_id> — poll a server-side image build; dumps last 2KB of logs on FAILED
• wait_for_replacement.py <workload_id> — poll a rolling replacement; handles the 404-when-cleared case
• check_limits.py — print the user's effective org-set scaling limits via /account/info/

Deeper docs in references/

SKILL.md is the operational core; occasional detail lives in references/:

• status-vocabulary.md — workload + proton status enums and transitions
• common-error-patterns.md — CrashLoopBackOff / ImagePullBackOff / OOMKilled / probe / exec-format / pending
• schema-reference.md — schemas to look up, credential-type→key maps, public-spec path quirks
• lifecycle-flows.md — artifact draft→lock→prod rules, replacement preconditions, redeploy matrix, imageUri gotchas
• code-to-workload.md — deploy from source: dr CLI, codeRef, Execution Environments, iterate-rebuild loop
• web-uis-behind-the-edge.md — browser-facing web app through the endpoint: prefix stripping, auth gate, Authorization hijack, shim, CSRF, WebSockets

OpenAPI spec is source of truth

At ${DATAROBOT_ENDPOINT}/openapi.yaml. ~5 MB — never dump it whole. Save once, then slice with yq (or print() only the specific key in Python):

curl -sS "${DATAROBOT_ENDPOINT}/openapi.yaml" -o /tmp/wapi-spec.yaml
yq '.components.schemas.CreateWorkloadRequest' /tmp/wapi-spec.yaml
yq '.components.schemas | keys | .[]' /tmp/wapi-spec.yaml | grep -i workload   # discover

All workload paths are keyed with the /api/v2/ prefix — see references/schema-reference.md.

1. Create / configure / scale

Run a container as a workload (the 90% case)

# spec.yaml — JSON also accepted; spec is sent verbatim
name: my-api-service
importance: low
artifact:
  name: my-api-service-artifact
  spec:
    type: service
    containerGroups:
      - name: default
        containers:
          - name: main
            imageUri: ghcr.io/org/my-app:latest
            port: 8000
            primary: true
            readinessProbe: {path: /readyz, port: 8000, initialDelaySeconds: 10}
            livenessProbe: {path: /healthz, port: 8000, initialDelaySeconds: 30}
runtime:
  containerGroups:
    - name: default          # must match artifact.spec.containerGroups[].name (above)
      replicaCount: 1
      containers:
        - name: main
          resourceAllocation: {cpu: 1, memory: "512MB"}

dr workload create --spec-file spec.yaml         # v0.2.74+; 4xx: 400=schema/limit, 403=cap (run check_limits.py), 409=name conflict
dr workload get <workload_id>                    # or `dr workload status` — poll until status=running

Lifecycle one-liners (v0.2.74+): dr workload {stop|start|delete|endpoint|list} <id>.

Raw fallback when CLI unavailable: httpx.post(f"{base}/workloads/", headers=headers, json=spec) + r.raise_for_status() + r.json()["id"]. Then python scripts/wait_for_running.py <workload_id>.

Critical gotchas:

• importance: low/moderate/high/critical; type: service (default) or nim. Exactly one container per group has primary: true.
• cpu is cores (float OK). memory accepts decimal string ("512MB", units B/KB/MB/GB) or byte integer; Kubernetes binary suffixes (Mi/Gi) NOT supported.
• port MUST be >= 1024. The container must actually listen on it (set via image env vars or entrypoint).
• Image must include a linux/amd64 manifest. Apple Silicon defaults to ARM64 and crash-loops with exec format error. Build with docker buildx build --platform linux/amd64,linux/arm64 -t <ref> --push ..
• Status lifecycle: submitted → provisioning → launching → running (happy path); updating during rolling redeploys; errored recoverable; failed/terminated unrecoverable. Full table in references/status-vocabulary.md.

Serving a browser-facing web UI through the endpoint

If the container serves a web app (UI + its own backend/API/WebSocket) opened in a browser via dr workload endpoint <id> (not a headless service), the DataRobot edge gateway serves it under a path prefix and: strips the prefix inbound (no outbound rewrite — the app must be sub-path aware); is the auth gate (DataRobot login required) and hijacks the Authorization header (→ 401 {"message":"Invalid API key"}, never reaching the container); passes WebSockets through. Winning pattern: set the app's base-path to the prefix + re-add it inbound (derive it from the injected WORKLOAD_ID), disable the app's own auth (trust the edge), disable CSRF, probe an unauthenticated path. Full guidance, shim code, and per-symptom diagnostics: references/web-uis-behind-the-edge.md.

"Update the workload" disambiguation

Replicas, resources, autoscaling

PATCH /workloads/{wid}/settings/ with full body shape — use exactly one of replicaCount or autoscaling. Read settings first via GET /workloads/{wid}/settings/, then PATCH back:

httpx.patch(
    f"{base}/workloads/{wid}/settings/",
    headers=headers,
    json={
        "runtime": {
            "containerGroups": [
                {
                    "name": "default",
                    "replicaCount": 3,
                    "containers": [
                        {
                            "name": "main",
                            "resourceAllocation": {"cpu": 2, "memory": "1GB"},
                        }
                    ],
                    # OR: "autoscaling": {"enabled": True, "policies": [{
                    #       "scalingMetric": "cpuAverageUtilization",
                    #       "target": 70, "minCount": 1, "maxCount": 10}]}
                }
            ]
        }
    },
)

Valid scalingMetric values: cpuAverageUtilization, httpRequestsConcurrency, gpuCacheUtilization, gpuRequestQueueDepth, or a custom NIM metric. Settings updates are rolling; zero-downtime only with replicaCount >= 2 (or autoscaling minCount >= 2).

Org-set scaling limits — check before scaling

Two admin-set caps: maxConcurrentWorkloads and maxWorkloadReplicas. Value 0 = unlimited; users can't change them. Read via GET /account/info/ — response includes {"limits": {"maxConcurrentWorkloads": N, "maxWorkloadReplicas": M}} (or python scripts/check_limits.py). The spec's /users/{uid}/ and /organizations/{id}/ paths require Admin API access. Exceeding either limit returns HTTP 403 with {"detail": "Requested replicas (N) exceeds the maximum allowed (M)."} — check limits first, then propose the max allowed or flag that admin help is needed.

GPU type / VRAM — set via compute bundle, not direct

resourceAllocation only accepts cpu, memory, gpu (count). There is NO gpuType or gpuMemory field. To target a GPU model / VRAM size: GET /mlops/compute/bundles/ lists bundles (cpu.small, gpu.l4.small, gpu.a10g.medium); pass via "resourceBundles": ["gpu.l4.small"] (a list, but exactly ONE bundle allowed) under the container group. When a bundle is set, CPU/memory in resourceAllocation are ignored — the bundle defines them.

Credential injection — never hardcode secrets

DataRobot credentials are stored centrally and injected into environmentVars by reference:

"environmentVars": [
    {"name": "PLAIN_VAR", "value": "literal-value"},
    {"source": "dr-credential", "name": "AWS_ACCESS_KEY_ID",
     "drCredentialId": "<credential-id>", "key": "awsAccessKeyId"},
]

Workflow: GET /credentials/?limit=50 → note the credential's credentialType → look up the valid key field names for that type in references/schema-reference.md (covers s3, basic, api_token, bearer, oauth, gcp, azure_*, databricks_*, snowflake_*, …).

Create from an existing artifact

Provide artifactId instead of the inline artifact block. The containerGroups[].name and containers[].name in runtime must match what the artifact defines.

2. Diagnose — workload is stuck, errored, or crash-looping

One command for the full diagnosis

python scripts/diagnose_workload.py <workload_id>

Runs all 5 steps below, prints a structured report (status / logTail signals / flagged events / proton K8s detail / evidence / recommended next step / console URL). --json for machine-readable. If Evidence is empty, pull application logs via section 3 — don't guess from status alone.

The 5-step flow

The script encapsulates this; use the model below for ambiguous output or one-off calls.

1. GET /workloads/{id}/ — status, statusDetails.logTail (~30 lines; scan for error/exception/traceback/killed/permission denied/connection refused), statusDetails.conditions. Guard statusDetails — it's null during submitted/provisioning.
2. GET /workloads/{id}/events/ — flag type: Warning or reason with Failed/Error/Kill/OOM; the last Warning before errored is usually the trigger.
3. GET /workloads/{id}/protons/ — pick role: "active" (or the candidate during a rolling replacement; else newest createdAt).
4. GET /workloads/{id}/protons/{pid}/statusDetails/ — 204 while initializing (not an error). Read replicas[*].containers[*].status+restartCount → replicas[*].conditions[*] (any value:false) → overallStatus.summary.
5. Application logs — section 3.

Common patterns (CrashLoopBackOff, ImagePullBackOff, OOMKilled, probe/pending, exec format error) and fixes: references/common-error-patterns.md.

Reporting findings

Workload {id} — Diagnosis
- Status: {current}
- Root cause: {one sentence}
- Evidence: {the specific logTail line, condition, container reason, or event}
- Recommended fix: {actionable next step — section 1 (settings), section 4 (artifact), or app code}
- Console: https://app.datarobot.com/console-nextgen/workloads/{id}/overview

3. Observe — logs, traces, metrics, service stats

Always check r.status_code before .json(): 401 = bad token; 404 = workload not found; 429 = rate limited (exponential backoff). All list endpoints accept limit + offset.

Logs

dr workload logs <wid> --level error --limit 100   # v0.2.74+; --follow streams; --output-format json

--level is an EXACT severity match (not a threshold). For substring filtering on the message body, or proton-scoped logs (find proton IDs in section 2), drop to REST — dr workload logs doesn't expose those filters:

r = httpx.get(
    f"{base}/otel/workload/{wid}/logs/",
    headers=headers,
    params=[
        ("searchKeys", "proton_id"),
        ("searchValues", pid),
        ("searchKeys", "level"),
        ("searchValues", "error"),
    ],
)

searchKeys / searchValues are positional parallel lists — pass a list of tuples to httpx (dict can't repeat keys). includes=<substring> does case-sensitive substring filtering on the message body.

Traces

traces = httpx.get(f"{base}/otel/workload/{wid}/traces/", headers=headers).json()[
    "data"
]
# summary: traceId, rootSpanName, rootServiceName, duration (NANOSECONDS), spansCount, errorSpansCount
trace_id = next(
    (t["traceId"] for t in traces if t.get("errorSpansCount", 0) > 0),
    traces[0]["traceId"],
)
trace = httpx.get(
    f"{base}/otel/workload/{wid}/traces/{trace_id}/", headers=headers
).json()

duration is NANOSECONDS on summaries AND spans. Divide by 1,000,000 for ms before display. Empty data = app isn't instrumented; direct the user to wire up OTEL.

Metrics + service stats

Convert before display: bytes→MB (/1024**2), nanocores→cores (/1_000_000), percentage already %.

stats = httpx.get(f"{base}/workloads/{wid}/stats/", headers=headers).json()
# {"period": {...}, "metrics": {totalRequests, serverErrors, userErrors, slowRequests,
#   responseTime, requestsPerMinute, concurrentRequests, *ErrorRate}}. /workloads/stats/ = aggregate.

Destructive: DELETE /workloads/{id}/stats/?metricName=<name> zeroes a metric's history — only on explicit request.

Presenting results

Logs: timestamp | level | message, ERROR/CRITICAL first. Traces: table sorted by errors desc then recency. Metrics: apply unit conversion before display. Service stats one-liner: "{totalRequests} requests, {totalErrorRate*100:.2f}% errors, {responseTime:.1f} ms avg, {requestsPerMinute} req/min." Empty data → say why (not running, not instrumented, empty window), don't just "no data".

4. Artifact lifecycle

An artifact is the immutable-after-lock definition of what a workload runs (image, port, env vars, probes). A workload is the running instance + its runtime (replicas, resources, autoscaling). Resources do NOT belong on the artifact.

Picking the right path

Find the running artifact (workload["artifactId"]), check artifact["status"]. A running workload does not auto-adopt a rebuild until you redeploy.

• Same draft (the C2W loop) — in-place change or rebuild. PATCH/rebuild the draft, then roll onto it with PATCH /workloads/{id}/settings/: re-send the runtime body (even unchanged values trigger a rolling 202 redeploy that re-reads the current spec + latest COMPLETED build). Zero-downtime at ≥2 replicas. (POST /replacement/ onto the same draft also works.)
• Different / locked artifact. POST /replacement/ onto the other artifact ID. Locked in-place edit: clone → PATCH clone → lock → replace onto the clone.

Lock: dr artifact lock <id> (= PATCH /artifacts/{id}/ {"status":"locked"}). Promote (POST /workloads/{wid}/promote/, 200) locks the running draft in place, no restart. Runtime-only changes (replicas/resources/autoscaling) → PATCH /settings/; a PATCH to the artifact doesn't affect live workloads until you redeploy.

Preconditions (status-match, same-artifact rule) and the full redeploy matrix: references/lifecycle-flows.md.

How does your image get to DataRobot?

The artifact's imageUri must point at a registry DataRobot can pull from (image-pull creds aren't accepted at workload creation yet). Two paths:

1. Bring your own image — public registry or one the admin pre-configured. docker buildx ... --platform linux/amd64, push, set imageUri. Default flow.
2. Code-to-Workload (C2W) — no local Docker / no public registry: dr artifact code init + sync, then dr artifact build create builds server-side, pushes to DataRobot's internal registry, and populates imageUri. Full flow in references/code-to-workload.md.

Poll builds with python scripts/wait_for_build.py <artifact_id> <build_id>; only drafts build. imageUri is build-managed — never PATCH it by hand (422 "not permitted on this cluster"), and never PATCH the spec mid-build (a whole-spec write clobbers the pending build image → redeploys the old one). Sequence spec edits before build create or after COMPLETED.

C2W is preview / feature-flagged — ENABLE_WORKLOAD_API_CONTAINERS=true (org) + DATAROBOT_CLI_FEATURE_WORKLOAD=true (client).

Rolling artifact replacement

httpx.post(
    f"{base}/workloads/{wid}/replacement/",
    headers=headers,
    json={
        "artifactId": new_artifact_id,
        "strategy": "rolling",  # only "rolling" supported
        "config": {"warmupDurationMinutes": 2, "keepOldVersionMinutes": 5},  # optional
        # "runtime": {...}  # optional; same shape as PATCH /settings/
    },
)

Monitor with python scripts/wait_for_replacement.py <workload_id>. Preconditions: status must match (draft↔draft / locked↔locked, else 400); same-artifact replacement 422s for locked but works for drafts — to roll the same draft without replacement use PATCH /settings/. Not idempotent (a second POST queues another swap); GET .../replacement/ 404 = none in progress; DELETE to cancel. Detail in references/lifecycle-flows.md.

Related skills

• datarobot-setup — install SDK, configure auth, set env vars
• datarobot-app-framework-cicd — declarative artifact + workload management via Pulumi and CI/CD
• datarobot-external-agent-monitoring — instrument arbitrary agent code with OTEL → DataRobot