derivation-python

Estuary derivation implemented as a Python class with async-generator transforms. Runtime is uv, types come from Pydantic models generated at flowctl generate time, and the whole thing is validated in pyright strict mode at publish.

Prereq: read derivation-basics first for concepts, project layout, stateless-vs-stateful, and the universal workflow.

Hard constraint: Python derivations run on private or BYOC data planes only — never on shared public planes. Reach out to your Estuary account manager if you're interested in a private or BYOC data plane.

Docs: concept page · walkthrough · upstream examples (stateful.flow.py, pipeline.flow.py).

When to reach for Python (vs SQL / TypeScript)

• ML / embeddings / tokenisers — model2vec, fastembed, transformers, tiktoken
• Async I/O — per-document await httpx.get(...), external APIs with bounded concurrency
• Heavy Python ecosystem deps — pandas, pyarrow, custom parsers
• Rich type safety — Pydantic validators; pyright-strict at publish

Reach for other skills when plain filter/transform fits SQL (derivation-filter-transform), the logic is reducible (derivation-aggregate-metrics), or you want stateful logic without a private DP (derivation-stateful-logic, SQLite).

How it works

You declare derive.using.python.module: <file>.flow.py. flowctl generate introspects your collection schemas and emits IDerivation, Document, and Request symbols you import. Your Derivation(IDerivation) class has one async generator method per transform — named as snake_case of the YAML transforms[].name. Each transform method consumes Request.Read<TransformName> (typed Pydantic read.doc) and yields Document(...) values. Dependencies in derive.using.python.dependencies are resolved by uv at publish time and at runtime. Optional state is persisted across restarts via __init__ / start_commit / reset lifecycle methods.

Canonical example — stateless

Rule-based topic classifier over a text stream. Real code from an audience-submission demo; minimal shape, 1:1 input→output.

Project layout

enrich/
├── flow.yaml
└── enrich.flow.py

flow.yaml

collections:
  acmeCo/classified/submissions:
    key: [/id]
    writeSchema:
      type: object
      required: [id, created_at, text, category]
      properties:
        id:         { type: string, format: uuid }
        created_at: { type: string, format: date-time }
        text:       { type: string }
        category:
          type: string
          enum: [ai-agents, fraud-risk, revenue, observability, other]
        keywords:
          type: array
          items: { type: string }
          default: []
    readSchema:
      allOf:
        - $ref: flow://write-schema
        - $ref: flow://inferred-schema

    derive:
      using:
        python:
          module: enrich.flow.py
      transforms:
        - name: enrich
          source: acmeCo/source/submissions
          shuffle: any

enrich.flow.py

"""Rule-based category bucketing + keyword extraction (stateless, 1:1)."""

import re
from collections.abc import AsyncIterator
from typing import Literal

from acmeCo.classified.submissions import IDerivation, Document, Request

Category = Literal["ai-agents", "fraud-risk", "revenue", "observability", "other"]

CATEGORIES: list[tuple[Category, set[str]]] = [
    ("ai-agents",    {"agent", "llm", "copilot", "rag", "prompt"}),
    ("fraud-risk",   {"fraud", "risk", "anomaly", "abuse"}),
    ("revenue",      {"revenue", "churn", "conversion", "upsell"}),
    ("observability", {"alert", "slo", "incident", "paging"}),
]
STOPWORDS = {"the", "and", "for", "that", "with", "you", "our"}
TOKEN_RE = re.compile(r"[a-zA-Z][a-zA-Z\-]{2,}")


def classify(tokens: set[str]) -> Category:
    for name, triggers in CATEGORIES:
        if tokens & triggers:
            return name
    return "other"


class Derivation(IDerivation):
    async def enrich(self, read: Request.ReadEnrich) -> AsyncIterator[Document]:
        src = read.doc
        tokens = TOKEN_RE.findall(src.text.lower())
        yield Document(
            id=src.id,
            created_at=src.created_at,
            text=src.text,
            category=classify(set(tokens)),
            keywords=[t for t in tokens if t not in STOPWORDS][:5],
        )

The import line's package path mirrors the collection name: acmeCo/classified/submissions → from acmeCo.classified.submissions import .... Slashes become dots; dashes in path segments become underscores.

Canonical example — stateful

Per-key running counts, persisted across restarts. Adapted from examples/derive-patterns/stateful.flow.py.

flow.yaml (excerpt)

derive:
  using:
    python:
      module: running_totals.flow.py
  transforms:
    - name: fromInts
      source: acmeCo/source/ints
      shuffle: { key: [/Key] }   # keyed shuffle for stateful work

Keyed shuffle is non-negotiable: if two docs for the same /Key land on different shards, each shard sees empty state and counts never accumulate. This is the same rule as derivation-stateful-logic.

running_totals.flow.py

from collections.abc import AsyncIterator
from pydantic import BaseModel, Field
from acmeCo.processed.running_totals import IDerivation, Document, Request, Response


class State(BaseModel):
    class KeyState(BaseModel):
        count: int
        sum: int
    keys: dict[str, KeyState] = Field(default_factory=dict)


class Derivation(IDerivation):
    def __init__(self, open: Request.Open):
        super().__init__(open)
        self.state = State(**open.state)         # rehydrate from prior transaction
        self.touched = State()                   # keys changed this transaction

    async def from_ints(self, read: Request.ReadFromInts) -> AsyncIterator[Document]:
        ks = self.state.keys.setdefault(read.doc.Key, State.KeyState(count=0, sum=0))
        ks.count += 1
        ks.sum += read.doc.Int
        self.touched.keys[read.doc.Key] = ks
        yield Document(Key=read.doc.Key, Count=ks.count, Sum=ks.sum)

    def start_commit(self, _: Request.StartCommit) -> Response.StartedCommit:
        updated = self.touched.model_dump()
        self.touched = State()
        return Response.StartedCommit(
            state=Response.StartedCommit.State(updated=updated, merge_patch=True),
        )

    async def reset(self):
        self.state = State()
        self.touched = State()

Lifecycle:

• __init__(self, open) runs once per task shard. open.state is the previously-persisted JSON (empty dict on first run).
• start_commit(self, _) runs at the end of each transaction. merge_patch=True persists only the keys that changed; False replaces the full state. Merge patch is strictly better for large state.
• reset(self) runs between catalog tests so they don't carry state between cases.

Full source: https://github.com/estuary/flow/blob/master/examples/derive-patterns/stateful.flow.py

Workflow

# 1. Pull the source collection locally
flowctl --profile <profile> catalog pull-specs --name acmeCo/source/submissions

# 2. Add the derive stanza to flow.yaml (as above)

# 3. Generate Pydantic stubs, pyproject.toml, pyrightconfig.json
flowctl --profile <profile> generate --source flow.yaml

# 4. Implement the class in <name>.flow.py

# 5. Preview locally against real source data
flowctl --profile <profile> preview --source flow.yaml \
  --name acmeCo/classified/submissions --timeout 30s

# 6. Publish to a PRIVATE data plane (required the first time for a new prefix)
flowctl --profile <profile> catalog publish --source flow.yaml \
  --init-data-plane ops/dp/private/estuary/aws-us-east-1-c1 \
  --auto-approve

flowctl generate creates flow_generated/python/<tenant>/.../__init__.py with typed IDerivation, Document, Request, Response, plus a skeleton .flow.py if absent. Re-running generate does not overwrite your implementation — only the flow_generated/ tree.

Dependencies

Declared under derive.using.python.dependencies as a map of package → version specifier (PEP 440). Installed by uv on container start — which means at publish, preview, and flowctl catalog test. A slow or failing install breaks all three.

derive:
  using:
    python:
      module: embed.flow.py
      dependencies:
        model2vec: ">=0.3.0"
        httpx: ">=0.27.0"
        numpy: ">=2.0"

pydantic >=2.0 and pyright >=1.1 are auto-included — don't list them.

Container environment

• Python version is pinned to 3.14 at the container level (ghcr.io/astral-sh/uv:python3.14-trixie-slim). You don't choose it — whatever's in the container runs. In your pyproject.toml, use requires-python = ">=3.12" (or >=3.14); narrower upper bounds like <3.13 cause uv to fail resolution.
• No build toolchain. The slim container has no rustc, gcc, or make. uv falls back to source builds when wheels aren't available for cp314-linux_x86_64 — that's where dependency selection fails.

What works

Pure-Python packages, numpy-family, and anything with pre-built cp314 wheels for Linux x86_64. Before adding a dep, check on PyPI that a wheel exists for cp314 (or that the package is pure-Python).

What fails

Packages that require native build tools or hefty downloads. Concrete examples from production use:

• sentence-transformers — pulls ~1.5 GB CUDA torch as a transitive dep; install times out under the container's network budget.
• fastembed — depends on py-rust-stemmers, which has no cp314 wheel and needs rustc to build from source.
• Anything declaring setuptools >=78.0.0 dependencies with dash-separated keys in setup.cfg — pin setuptools<78 alongside if you hit Invalid dash-separated key 'description-file'.
• Packages with wheels only up to cp312 or cp313 — uv will attempt a source build and fail.

For embeddings specifically, model2vec (pure-numpy lookup via distilled static vectors) is the path of least resistance — ships pre-built wheels, ~30 MB model download, no native deps.

Debugging install failures

If publish/preview/test fails before your code runs, the install step is the most likely culprit. Reproduce locally with uv pip install --python 3.14 <pkg> against a fresh venv to check wheel availability before committing to a dependency.

Async patterns

Every transform is an async generator. You can await I/O freely, but per-doc serial awaits limit throughput — for API calls, pipeline with bounded concurrency (full example: https://github.com/estuary/flow/blob/master/examples/derive-patterns/pipeline.flow.py):

import asyncio
import httpx
from typing import AsyncIterator
# IDerivation, Document, Request imported from your generated module

class Derivation(IDerivation):
    MAX_CONCURRENT = 10

    def __init__(self, open: Request.Open):
        super().__init__(open)
        self.client = httpx.AsyncClient()   # shared across calls; reused within shard lifetime
        self.pending: set[asyncio.Task[Document]] = set()

    async def _enrich(self, src) -> Document:
        # Per-doc work — e.g. await an external API for each input doc.
        resp = await self.client.get(f"https://api.example.com/lookup/{src.id}")
        return Document(id=src.id, **resp.json())

    async def enrich(self, read: Request.ReadEnrich) -> AsyncIterator[Document]:
        if len(self.pending) >= self.MAX_CONCURRENT:
            done, self.pending = await asyncio.wait(self.pending, return_when=asyncio.FIRST_COMPLETED)
            for t in done:
                yield await t
        self.pending.add(asyncio.create_task(self._enrich(read.doc)))

    async def flush(self) -> AsyncIterator[Document]:
        # Drain remaining tasks at transaction close.
        for doc in await asyncio.gather(*self.pending):
            yield doc
        self.pending.clear()

httpx.AsyncClient is hoisted onto self so a single connection pool is reused for the shard's lifetime — constructing a new client per call leaks connections and defeats keep-alive.

flush() is a lifecycle hook the runtime calls at the end of each transaction, separate from your transform methods. Use it when you've buffered pending work.

Runtime container constraints

• HOME=/nonexistent — huggingface-hub crashes with PermissionError trying to write ~/.cache/huggingface/. Set HF_HOME + XDG_CACHE_HOME before importing the library:

  import os
  os.environ.setdefault("HF_HOME", "/tmp/hf_cache")
  os.environ.setdefault("HF_HUB_CACHE", "/tmp/hf_cache/hub")
  os.environ.setdefault("XDG_CACHE_HOME", "/tmp/xdg_cache")
  from model2vec import StaticModel   # safe to import now
  

• /tmp is ephemeral — model re-downloads on every shard cold start. Fine for 30 MB; plan otherwise for 500 MB+.

• Persistent state lives in RocksDB on the reactor's /mnt/local and is shared with other tasks. No enforced cap — runaway per-task state can cause reactor-wide disk pressure, so self-police state growth.

Testing

Local — flowctl preview --fixture

flowctl --profile <profile> preview --source flow.yaml \
  --name acmeCo/classified/submissions \
  --fixture fixture.jsonl --timeout 30s

fixture.jsonl:

["acmeCo/source/submissions", {"id":"1","created_at":"2026-04-24T10:00:00Z","text":"Fraud detection that sees transactions in seconds"}]
["acmeCo/source/submissions", {"id":"2","created_at":"2026-04-24T10:01:00Z","text":"LLM agent that schedules meetings"}]
{"commit": true}

Preview runs the derivation locally (no private DP needed for preview itself), but does require a Docker daemon — Python derivations execute inside ghcr.io/estuary/derive-python:dev. Same for flowctl catalog test. Set DOCKER_CLI=podman if you prefer podman. Stateful derivations build state within the session — use --sessions to exercise restart behavior.

Gotchas specific to local preview:

• Preview can fail with No space left on device at /tmp/.tmpXXX — local disk, not reactor. Clean /tmp and retry.
• Preview runs everything in-process — shard-splitting semantics don't match production. Good enough for correctness checks, not for load testing.

Spec-defined tests — flowctl catalog test

Same tests: block syntax as other derivation types — see derivation-basics. reset(self) is called between test cases.

Caveat for Python derivations: flowctl catalog test validates the spec but times out waiting for the Python container to become ready (shard-readiness window is too short for Python bootstrap). For runtime validation, publish to your private data plane and inspect the derived collection.

<!-- TODO: add a concrete `tests:` block for the `enrich` canonical example — ingest a few `acmeCo/source/submissions` docs covering each category branch + the "other" fallback, then `verify` the resulting `acmeCo/classified/submissions` documents. Include a stateful test for the `running_totals` example exercising `reset()` between cases. -->

Gotchas specific to Python

• Private/BYOC only. Publishing a Python derivation to a public data plane fails at container start with "Python derivations may only run in private data-planes". Use --init-data-plane ops/dp/private/... (or set the storage mapping) to land on a private plane.
• Transform method name = snake_case of the YAML name. name: fromOrders → async def from_orders. Mismatch raises TypeError: Can't instantiate abstract class Derivation with abstract method <name> at startup — the generated IDerivation base declares each transform as @abstractmethod, so a missing override blocks instantiation.
• async def + yield is mandatory. Regular def or missing yield isn't registered as an async generator — no documents flow. For state-only transforms keep the stub's if False: yield line.
• Import from your own generated module for pyright coverage — from acmeCo.my.collection import IDerivation, Document, Request. Path-based imports work but lose type safety.
• flow_generated/ is regenerated every flowctl generate — never edit. Edit only .flow.py and the collection YAML.
• merge_patch=True is almost always right for start_commit. False rewrites full state JSON every commit.
• Model re-downloads on every shard cold start. Pre-warm with a dummy ingest before anything live.
• Cross-plane source reads (e.g., derivation on private reading a collection on public) work between non-legacy data planes as of 2026. The documented exception is firewalled BYOC planes, where cross-plane reads still don't work. Confirm specifics with your account manager for customer deployments.

Related

• derivation-basics — prerequisite reading
• derivation-stateful-logic — SQLite alternative for stateful work on public planes
• derivation-aggregate-metrics / derivation-windowing — reductions that don't need procedural state
• Python concept page
• Python tutorial
• Upstream examples
• Private deployments overview