interlace

Python/SQL-first data platform: transformation, orchestration, and durable streaming — one process.

interlace is an independent, MIT-licensed alternative to dbt/SQLMesh that also replaces the orchestrator (no Airflow) and the ingestion layer (Cloudflare-Pipelines-style durable streams). Models are .sql files or Python functions; state is versioned snapshots with virtual environments and a terraform-style plan/apply; everything runs in a single daemon on DuckDB + DuckLake by default.

Status: 1.0. Requires Python 3.12+. The package is published to PyPI as interlaced; the import name and CLI are interlace.

pip install 'interlaced[service]'   # the CLI + daemon; core CLI only: pip install interlaced
# more extras: [adbc] postgres/redshift · [spark] · [polars] · [all]

Sixty seconds

interlace init my-project && cd my-project
interlace plan            # terraform-style preview: added / breaking / non-breaking / reuse
interlace apply           # build changed models, run checks, promote the environment
interlace serve           # the daemon: web UI (/ui) + HTTP API + scheduler + streams, one process

Every model builds into a fingerprinted physical table (interlace__main.orders__a1b2c3); environments are views over those tables, so promotion and rollback are atomic view swaps and a dev environment reuses prod's tables for free. Production is the unprefixed namespace — consumers query main.orders; sandboxes are prefixed (dev__main.orders). Commands default to prod; pass --env dev while developing.

Models

SQL — a file per model; upstreams referenced by model name, dependencies inferred by parsing (sqlglot), config in a leading comment block:

/* interlace:
  strategy: scd
  key: customer_id
  schedule: {cron: "0 * * * *"}
  checks:
    - not_null: customer_id
    - unique: customer_id
*/
SELECT customer_id, name, tier FROM raw_customers

Python — functions whose parameters name their upstreams; data crosses as Arrow (never pandas), streamed with bounded memory:

from interlace import model

@model(strategy="merge", key="order_id", cursor="updated_at")
def orders(cursor, this):
    """Incremental API extract: `cursor` is max(updated_at) already in the
    warehouse (None on first run); `this` is the previous materialisation."""
    rows = fetch_orders(since=cursor)          # your code
    return pyarrow.Table.from_pylist(rows)     # or RecordBatchReader / generator of batches

Strategies: replace, view, ephemeral (CTE-inlined), merge (upsert), full_merge (full-state source applied as a minimal diff), incremental_by_time (windowed, interval-ledger backfill/catchup), scd (history with validity windows).

Plan / apply

$ interlace plan
 Model         Change    Category      Build
 orders        modified  non_breaking  rebuild
 order_stats   modified  non_breaking  reuse      <- output provably identical: not rebuilt

• Changes classify breaking / non-breaking / forward-only; a plan with breaking changes refuses to apply without --force. Downstream models whose output is provably identical (column-pruned impact analysis) reuse their existing tables instead of rebuilding — an improvement over model-granular invalidation.
• apply --forward-only lets history-keeping models (scd2/merge/incremental) survive a definition change: the existing table is copied to the new version, the new logic applies to the copy going forward, and checks gate before views move.
• Checks gate promotion: 10 built-in types (not_null, unique, accepted_values, row_count, freshness, expression, relationships, pattern, range, sql) plus @check Python functions — an error-severity failure blocks before the environment view moves. interlace checks run re-runs them ad hoc against any environment's promoted tables.
• --select state:modified scopes a plan/apply to models whose fingerprint drifted from the target environment, plus everything downstream — the CI diff, one flag. interlace impact model.column shows the column-level blast radius of a change.
• interlace env rollback repoints an environment's views at any earlier promotion — nothing rebuilds, the views move; every apply records a generation, so a bad deploy reverts in one command. interlace gc removes snapshots no environment references (reference-aware: tables shared through reuse survive).

Streaming

Declare a stream; POST to it; rows are durable (SQLite WAL log) before the 200, deduplicated by idempotency key, and materialized exactly-once into streams.<name> — a micro-batch flusher commits the data and the watermark in one warehouse transaction, and SQL models just read the table. A flush triggers the models that consume the stream.

from interlace import stream

@stream("orders", schema={"order_id": "string", "total": "double"},
        idempotency_key="order_id", retention="7d", on_schema_drift="evolve")
def orders(event): ...

curl -X POST localhost:8000/streams/orders -d '{"order_id": "o1", "total": 49.5}'

Schema drift is yours to choose: reject (400), evolve (new columns appear), or quarantine (bad events divert to <stream>__quarantine). When the warehouse falls behind, publishes get 429 backpressure instead of unbounded backlog.

Reverse ETL

Attach external databases and deliver model results into them — the live table is never dropped, keyed modes reuse the same merge strategies:

# interlace.yaml
attach:
  crm: "postgres:host=... dbname=crm"

/* interlace: {export: {to: table, target: crm.public.accounts, mode: merge, key: id}} */
SELECT id, tier, lifetime_value FROM account_summary

File exports (to: parquet|csv|json) work the same way. Sinks are environment-gated: by default the side effect fires only from prod — a dev apply never writes to a live external table (opt in with environments: [dev, prod]).

Multi-engine

Models run on named engines: DuckDB/DuckLake by default, Postgres natively over ADBC (pip install 'interlaced[adbc]'), Spark (beta, [spark] extra), plus alpha adapters for MotherDuck, Redshift, Snowflake and BigQuery (wired and dialect-correct, not yet run against a live account), with per-model pinning:

engines:
  pg: {type: postgres, database: "${PG_DSN}"}

/* interlace: {engine: pg, strategy: merge, key: id} */

Strategies execute inside the pinned engine (no DuckDB middleman); cross-engine dependencies appear as explicit transfer lines in the plan and move as Arrow (or a federated ATTACH fast lane when possible). Contract: docs/architecture/MULTI_ENGINE.md.

The daemon

interlace serve runs everything in one process:

• the web UI at /ui (in-package, zero build step) — ten views: overview, lineage canvas with column-level tracing, models, plan/apply with SQL diffs, live runs, query console, streams, checks, environments, and system — live over SSE;
• the HTTP API (Litestar + msgspec, OpenAPI at /schema/scalar) with the same surface as the CLI: plan/apply, runs, checks, streams, engines, schedules, lineage, query, gc;
• the scheduler: cron/interval triggers enqueue onto a durable run queue (leases, retries, cooperative cancellation — interlace cancel <id> or POST /runs/{id}/cancel);
• stream ingestion and retention.

Scoped API keys (interlace apikey create ci --scope read) lock it down; a durable event log backs GET /events/stream (SSE with Last-Event-ID replay).

Add --quack quack:localhost:4213 to serve the warehouse itself over DuckDB's quack protocol — other processes (CLI runs, ad-hoc DuckDB clients) then share it concurrently by setting database: quack:localhost:4213.

Architecture in five lines

• The IR is a sqlglot AST; the wire format is an Arrow RecordBatchReader; strategies are AST builders and dialect appears only at transpile().
• Storage defaults to DuckLake (Parquet + SQL catalog) opened as DuckDB's primary database.
• Control plane (snapshots, intervals, queue, events, keys) is SQLite WAL; Postgres is the scale-out swap.
• Streams live in their own durable log; the materializer commits data + watermark in one warehouse transaction — exactly-once without distributed coordination.
• No Jinja, no pandas in core, no external orchestrator.

The full design rationale lives in docs/architecture/architecture.md.

Development

Toolchain is pinned with proto, tasks run via moon, uv owns the virtualenv:

proto install
moon run interlace:sync      # install deps
moon run interlace:test      # 350+ tests
moon run interlace:check     # black + ruff (CI equivalent)
moon run interlace:typecheck # mypy

MIT licensed.