Incremental Models

Incremental models process only new or changed data instead of rebuilding the entire table. In the default sequential path, SQLBuild works out where to resume from the current target and input relations with no separate checkpoint store. A retry recomputes its interval from current warehouse state rather than reusing the exact interval of a failed attempt. Opt-in concurrent microbatching adds immutable coordination facts as described below.

Strategies

append

Inserts new rows without modifying existing data. Optionally uses a cursor (read from the target table’s highest value) to avoid reprocessing the full source on every run.

MODEL (
  materialized incremental,
  incremental_strategy append,
  cursor created_at,
  cursor_type timestamp,
  cursor_grain second,
  append_cursor_inclusive true,
);

SELECT id, customer_id, created_at
FROM __source("raw_events")

When append_cursor_inclusive is true (the default), the lower bound uses >=, which may duplicate the boundary row but avoids missing late-arriving data with the same cursor value. Set to false for an exclusive (>) lower bound if your cursor values are guaranteed unique. Append without a cursor is also valid. The model simply inserts all rows from the source query on every run.

delete_insert

Deletes rows in the cursor range, then inserts the new delta. Requires either cursor or unique_key.

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  unique_key [order_id],
  cursor order_id,
  cursor_type integer,
);

SELECT order_id, customer_id, order_status, ordered_at, line_total_cents
FROM __ref("fact_orders")

With a cursor, delete_insert removes rows where the cursor column falls within the replay window, then inserts the new delta. With a unique_key only, it deletes matching rows by key before inserting.

merge

Upserts rows using a unique key. Matched rows are updated; unmatched rows are inserted.

MODEL (
  materialized incremental,
  incremental_strategy merge,
  unique_key [customer_id],
  cursor last_ordered_at,
  cursor_type timestamp,
  cursor_grain second,
  cursor_inputs (
    fact_orders ordered_at,
  ),
);

SELECT
  customer_id,
  MAX(ordered_at) AS last_ordered_at,
  COUNT(*) AS total_orders,
  SUM(line_total_cents) AS total_revenue_cents
FROM __ref("fact_orders")
GROUP BY customer_id

merge always requires unique_key. The cursor controls which upstream rows are scanned; the unique key determines how they’re matched against the target.

Full-refresh overrides

Incremental models can override command-level full-refresh behavior with full_refresh:

Model setting	Normal command	Command with `--full-refresh`
omitted	incremental	full refresh
`full_refresh false`	incremental	incremental
`full_refresh true`	full refresh	full refresh

Use full_refresh false for models that must retain their normal cursor or microbatch execution even when a broader job requests full refresh:

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  full_refresh false,
  cursor event_date,
  cursor_type timestamp,
  cursor_grain day,
);

The override is evaluated per model, so one selection can contain incrementally executed opt-outs and full-refreshed models. It controls execution mode rather than acting as a safety rejection: an opted-out model does not abort the rest of the build. This does not skip initial loading. If the destination relation does not exist, an incremental or microbatch model still builds the history required by its cursor policy. Cloning an existing destination before the build can provide a current watermark and avoid a first-run historical replay.

Cursors

Cursors define the incremental replay boundary. SQLBuild queries MAX(cursor) from the target table and MIN/MAX from upstream inputs to compute the replay window automatically. Observed maxima are inclusive warehouse values; SQLBuild advances them once to produce an effective exclusive end bound.

Field	Description
`cursor`	Output column used to track incremental position
`cursor_type`	`timestamp` or `integer`
`cursor_grain`	Time grain for timestamp cursors: `second`, `minute`, `hour`, `day`, `month`, `year`
`cursor_start`	Lower bound floor; the cursor will never replay before this value
`cursor_inputs`	Map of upstream ref/source names to their cursor columns

cursor_inputs

When a model references multiple upstream inputs, cursor_inputs is required to tell SQLBuild which column on each input carries the cursor:

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_hour,
  cursor_type timestamp,
  cursor_grain hour,
  cursor_inputs (
    fact_orders ordered_at,
  ),
);

SQLBuild uses these to determine the replay window. With multiple listed inputs, the end is the conservative common watermark: the minimum of their maxima. This prevents a faster input from advancing the model beyond data available from a slower input. On a first build, SQLBuild derives the interval from the declared cursor inputs and cursor policy. If it cannot establish a valid interval, the build fails before mutating the destination.

Cursor bounds in model SQL

Cursor-based incremental models can read their effective interval with zero-argument intrinsics:

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_hour,
  cursor_type timestamp,
  cursor_grain hour,
  cursor_inputs (
    fact_orders ordered_at,
  ),
);

SELECT customer_id, ordered_at AS activity_hour
FROM __ref("fact_orders")
WHERE ordered_at >= __cursor_start()
  AND ordered_at < __cursor_end()

__cursor_start() is the effective inclusive start and __cursor_end() is the effective exclusive end after cursor floors, lookback, replay policy, and command-line overrides have been applied. The intrinsics accept no arguments and are only valid in built-in cursor incremental model query SQL. They are rejected in functions, hooks, audits, SQL tests and scenarios, source expressions, non-incremental or cursorless models, custom materializations, and non-microbatch full refreshes. In microbatch mode, the intrinsics resolve to each batch’s concrete bounds. A microbatch full refresh discovers its range from current inputs while ignoring the old destination watermark.

Listed inputs bound the window; unlisted inputs do not

Only the inputs you list in cursor_inputs bound the replay window. This is an explicit choice, and it has two consequences worth understanding:

Listed inputs drive the window. Their new data advances the MAX, which is what tells SQLBuild how far to reprocess and which rows of the target to rewrite.
Unlisted inputs are read in full. SQLBuild does not add a cursor filter to them, and they do not bound the window. This is correct for lookup or dimension tables that have no meaningful cursor column: you do not list them, and SQLBuild reads them whole rather than trying to filter on a column that may not exist.

The implication for delete_insert and merge: the target rows that get rewritten are the ones whose cursor falls inside the window derived from the listed inputs. If an unlisted input changes in a way that should affect target rows outside that window, those rows are not rewritten on a normal incremental run. List every input whose new data should drive reprocessing; leave unlisted only the inputs you intend to read in full. To capture changes that fall outside the normal forward window, see Lookback for late-arriving data and Replay on change for model changes.

Lookback

Lookback extends the start of the replay window backwards to re-process recent data. The cursor is forward-moving, so use lookback to capture late-arriving or backfilled records that land just behind the current position:

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor event_date,
  cursor_type timestamp,
  cursor_grain day,
  lookback 3d,
);

With lookback 3d, the replay window starts 3 days before the normal cursor position, ensuring that any late-arriving data within that window is picked up.

Microbatch execution

For large incremental ranges, microbatch mode splits the replay window into configurable batches. Choose an explicit strategy:

Strategy	Window source
`watermark`	Declared input availability; supports timestamp and integer cursors
`rolling_window`	A timestamp window relative to the current run rather than input maxima

Each batch has its own audit cycle: create delta, run delta audits, apply DML, and clean up.

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_hour,
  cursor_type timestamp,
  cursor_grain hour,
  incremental_mode microbatch,
  microbatch_strategy watermark,
  cursor_watermark_mode all,
  cursor_inputs (
    fact_orders (column ordered_at, roles [filter, watermark]),
  ),
  batch_size 1d,
);

Without microbatch mode, the entire replay range is processed in one pass.

Batch size

batch_size controls the window size for each batch. For timestamp cursors, use duration strings like 1d, 6h, 1mo. For integer cursors, use an integer value. For watermark models, each cursor_inputs entry declares whether the input filters model SQL, contributes an availability watermark, or does both. cursor_watermark_mode all uses the conservative common watermark; any permits progress from any watermark input. A filter-only input does not claim that its intervals are available.

Concurrent batches

Sequential execution is the default and does not read or write _sqlbuild_microbatches. To opt into concurrent delete_insert batches, enable the project gate and choose a model limit:

[settings]
microbatch_concurrency = true

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  incremental_mode microbatch,
  microbatch_strategy watermark,
  batch_concurrency 4,
  -- cursor configuration omitted
);

Concurrent execution uses immutable requirement and completion facts to coordinate dependencies and reconcile retries. It does not update one mutable status row through planned/running/complete states. Increase concurrency deliberately because every active batch can consume warehouse work.

Watermark batch limits

Watermark microbatch models can declare what to do when their resolved range contains more batches than an ordinary run should process:

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  incremental_mode microbatch,
  microbatch_strategy watermark,
  cursor_watermark_mode all,
  cursor event_date,
  cursor_type timestamp,
  cursor_grain day,
  cursor_inputs (
    raw_events (column event_date, roles [filter, watermark]),
  ),
  batch_size 1d,
  microbatch_limit (
    max_batches 7,
    action cap_from_end,
  ),
);

Action	Behavior when the range exceeds `max_batches`
`error`	Fail before hooks or model mutation
`warn`	Emit a prominent warning and execute the full range
`cap_from_start`	Execute the earliest N aligned batches and defer later work
`cap_from_end`	Execute the latest N aligned batches ending at the resolved watermark and defer older work

A cap changes only the work selected for that invocation. Deferred batches are not recorded as complete. cap_from_end is useful for feeds where keeping the latest projection current is more important than catching up oldest-first; cap_from_start is the oldest-first catch-up policy. Downstream watermark models consume durable completion facts from capped upstream models. A target table’s physical minimum/maximum does not prove that deferred or intervening intervals were built; if completion evidence is unavailable, SQLBuild fails closed rather than inventing availability. The project can also set an outer safety policy:

[microbatches.limits]
max_batches = 100
action = "error" # or "warn"

Project limits support only error and warn; they never silently cap work. When a model has a nested microbatch_limit, the project limit checks the full resolved range first and the model policy then applies. --max-microbatches N is an invocation-wide, hard error ceiling and an explicit one-run authorization. It takes precedence over project and model limits, applies to models without a declared limit, and never inherits cap_from_start or cap_from_end. For example, passing a value large enough for an intentional backfill authorizes the full range instead of retaining the model’s ordinary-run cap. The legacy scalar max_microbatches model field remains supported as a fail/warn guard. New models should use the nested form when the action is part of the model’s execution policy.

Mixed-grain chains

When a downstream microbatch model reads from an upstream model with a coarser time grain, SQLBuild aligns the replay to the coarsest participating grain (the model’s own grain and its cursor-input grains). This happens on every run that resolves cursor bounds from upstream models, not only when something changes. It is independent of the replay_on_change cascade behavior described below. Alignment does two things: it floors the replay window edges to the coarsest grain, and it coarsens the batch size to that grain. For example, an hourly model downstream of a daily model processes in day-sized batches, so each batch lines up with a unit of upstream data that actually advances instead of producing empty or boundary-straddling windows:

-- Upstream: daily grain, 2d batches
MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_day,
  cursor_type timestamp,
  cursor_grain day,
  cursor_inputs (
    hourly_order_activity activity_hour,
  ),
  incremental_mode microbatch,
  batch_size 2d,
);

-- Downstream: hourly grain, but aligns to day automatically
MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_hour,
  cursor_type timestamp,
  cursor_grain hour,
  cursor_inputs (
    daily_activity_rollup activity_day,
  ),
  incremental_mode microbatch,
  batch_size 6h,
);

Replay on change

When a model’s version identity changes (query, config, upstream cascade, or any other change reason), replay_on_change is the explicit, per-model policy for how much data to reprocess. Reprocessing is a policy you set, not an automatic forced rebuild, so a definition change does not silently trigger a full rebuild of large downstream tables. You choose the cost per model:

Value	Effect
`forward`	Run the normal incremental delta from the cursor (default)
`full`	Full table rebuild
`bounded-14d`	Replay the last 14 days of data

The bounded duration supports d (days), h (hours), m (minutes), and s (seconds). For example: bounded-7d, bounded-24h, bounded-30m.

MODEL (
  materialized incremental,
  incremental_strategy delete_insert,
  cursor activity_hour,
  cursor_type timestamp,
  cursor_grain hour,
  replay_on_change full,
);

See Cascade propagation for how replay policies propagate through the DAG and how downstream models can override inherited replay behavior.

on_schema_change

Controls how schema differences are handled at execution time when the incremental delta has different columns than the target table:

Value	Effect
`append_new_columns`	Add new columns to the target table (default)
`sync_all_columns`	Add, drop, and alter columns to match the delta
`ignore`	Log and continue without schema changes
`fail`	Reject the build with an error

Getting Started

dbt Compatibility

Concepts

Advanced Concepts

Python Nodes

Virtual Environments (Alpha)

Integrations

CLI Reference

Incremental Models

Strategies

append

delete_insert

merge

Full-refresh overrides

Cursors

cursor_inputs

Cursor bounds in model SQL

Listed inputs bound the window; unlisted inputs do not

Lookback

Microbatch execution

Batch size

Concurrent batches

Watermark batch limits

Mixed-grain chains

Replay on change

on_schema_change

​Strategies

​append

​delete_insert

​merge

​Full-refresh overrides

​Cursors

​cursor_inputs

​Cursor bounds in model SQL

​Listed inputs bound the window; unlisted inputs do not

​Lookback

​Microbatch execution

​Batch size

​Concurrent batches

​Watermark batch limits

​Mixed-grain chains

​Replay on change

​on_schema_change

Strategies

append

delete_insert

merge

Full-refresh overrides

Cursors

cursor_inputs

Cursor bounds in model SQL

Listed inputs bound the window; unlisted inputs do not

Lookback

Microbatch execution

Batch size

Concurrent batches

Watermark batch limits

Mixed-grain chains

Replay on change

on_schema_change