Models

Each model is one SQL file with a MODEL() header and one outer SELECT. Models default to streaming tables: SQLTide creates a ClickHouse table and the materialized view that writes into it. A kind view model instead creates an ordinary terminal query view.

MODEL (
  engine "MergeTree()",
  order_by ["order_id", "_replay_partition", "_replay_offset"],
  partition_by "toYYYYMM(event_at)",
  ttl "event_at + INTERVAL 30 DAY",
  settings (
    index_granularity 8192,
  ),
  replay_anchor auto,
);

SELECT
  CAST(order_id AS String) AS order_id,
  amount_cents::UInt64 AS amount_cents,
  event_at::DateTime64(3) AS event_at,
  _replay_partition::Int32 AS _replay_partition,
  _replay_offset::Int64 AS _replay_offset,
  _replay_timestamp::DateTime64(3) AS _replay_timestamp
FROM __source("orders")

The header grammar is:

• fields are whitespace-separated key value entries
• lists use [...]
• nested mappings use (...)
• commas are optional separators and may trail the final entry
• values containing spaces must be quoted

YAML-like key: value and {...} mappings are not accepted. MODEL (); uses all defaults.

SQL contract

The outer statement must be one SELECT. Every outer projection must:

• be explicit rather than *
• have a unique alias
• declare an exact ClickHouse type with CAST(expr AS Type) AS alias or expr::Type AS alias

CTEs may contain stars, untyped expressions, and set operations. A top-level UNION, INTERSECT, or EXCEPT is rejected; wrap it in a CTE and finish with a typed projection.

order_by, partition_by, and ttl expressions may reference only output columns.

Header fields

Field	Default	Purpose
kind	table	table for streaming storage or view for an ordinary terminal view
relation_name	none	Exact warehouse relation name, overriding prefix-based naming
engine	"MergeTree()"	ClickHouse MergeTree-family table engine
order_by	["_replay_timestamp"]	Table sorting key expressions
partition_by	none	Partition expression
ttl	[defaults].model_ttl or none	ClickHouse TTL expression
retention	model/pipeline/project policy	Typed expiry mapping or false; see retention
settings	none	Table settings mapping
replay_anchor	auto	Use replay_anchor never to prevent this model becoming a replay anchor
replay_on_change	inherited	Virtual-environment change policy
bounded_replay_fallback	full	Virtual-environment fallback for unseedable models

replay_on_change and bounded_replay_fallback are rejected in direct mode. replay_anchor remains valid because direct rebuilds also need replay roots.

engine, order_by, partition_by, ttl, retention, settings, replay_anchor, replay_on_change, and bounded_replay_fallback are table-only and are rejected on views.

An explicit model ttl overrides [defaults].model_ttl. SQLTide validates the effective expression against the model’s output columns and includes it in table storage identity, desired state, and drift detection.

Driving references

A model has exactly one untyped driving input:

FROM __source("orders")

or:

FROM __ref("orders")

Source-root models use __source. Downstream models use __ref.

Additional joins and subqueries must declare their role:

FROM __ref("orders") AS orders
LEFT JOIN __ref("region_lookup", ref_type="reference") AS regions
  ON orders.region_code = regions.region_code

ref_type	Meaning
reference	A side dependency expected to be stable for replay
mutable	A side dependency whose current state may differ from historical processing-time state

Mutable references emit a replay warning and prevent the model from being a replay anchor. Both side-reference types participate in graph scope and ordering.

The single driving edge determines the streaming and replay path. Reference and mutable side edges both affect graph ordering and scope, but a mutable side edge also prevents the receiving model from becoming a replay anchor. A terminal view may read several graph relations for query-time joins, but it has no outgoing edge to another model.

Terminal views

A terminal view has no driving input and may read any number of sources or models:

MODEL (
  kind view,
  relation_name customer_orders,
);

SELECT
  orders.order_id::UInt64 AS order_id,
  payments.amount_cents::UInt64 AS amount_cents
FROM __ref("orders") AS orders
JOIN __ref("payments") AS payments USING (order_id)

View references do not use ref_type: every reference is an ordinary query dependency. Views do not participate in source inference, replay roots, watermarks, retention, or replay population. A view must have zero downstream model edges across the complete project graph, including cross-pipeline references. Other models cannot reference it, while tests and audits may target it.

Without relation_name, tables use the effective table_prefix and views use view_prefix. Pipeline naming overrides project naming, whose defaults are tbl__ and view__. Logical refs continue to use filename stems. kafka__, raw__, and mv__ are reserved internal prefixes.

Replay columns

Normalized replay columns are source-independent:

_replay_partition
_replay_offset
_replay_timestamp
_replay_landed_at
_replay_cursor

Managed Kafka landing exposes partition, offset, timestamp, and landed-at values. Adopted sources expose the roles declared in their column mapping. Preserve the columns required by the source mode through non-aggregate models when those outputs should remain eligible replay anchors.

Aggregate models do not need to expose post-aggregate replay columns. SQLTide places replay predicates on their input query instead.

Engines and duplicate delivery

Common engines include MergeTree, ReplacingMergeTree, SummingMergeTree, and AggregatingMergeTree, including replicated variants. Engine choice is part of replay correctness: the replay cutoff is inclusive, so authored SQL and storage engines must tolerate duplicate delivery where necessary.