dbt Incremental Model Strategies

By Michael Samuel Naeem · September 10, 2026 · 6 min read

Nightly full scan on a ten terabyte fact—materialization stayed table instead of merge incremental.

Strategy matrix

Append for immutable events; merge with unique_key for upserts; delete+insert for partition replaces.

A production team running dbt incremental models discovered that strategy matrix failures show up only when upstream dependencies shift traffic mix—staging load tests with uniform QPS missed the regression until Black Friday.

Runbook entry for strategy matrix: confirm blast radius (single namespace vs fleet-wide), identify last config change, roll back via documented single step, then capture metrics screenshots for postmortem—not ad-hoc dashboard hunting.

For dbt incremental models, instrument strategy matrix with low-cardinality metrics tied to user- visible outcomes: error rate, tail latency, freshness, or cost per successful operation—avoid paging on vanity gauges that never correlated with past incidents.

Game day scenario for strategy matrix: inject partial outage in staging quarterly, verify on-call can execute rollback in under fifteen minutes using only the linked runbook, update runbook with what actually broke.

Ownership for strategy matrix belongs in the service catalog with named rotation, last drill date, and known sharp edges—new engineers should deploy a safe canary within one week using that doc alone.

Change management for dbt incremental models: require peer review from someone outside the authoring team before production promotion—fresh eyes catch assumptions embedded in strategy matrix configs that authors no longer notice.

Capacity planning note: estimate peak QPS or job concurrency for strategy matrix, multiply by headroom factor one-point-five to two, compare against cloud quotas and license limits before launch week—not during the first outage.

Incremental predicates

is_incremental branch filters on watermark; include lookback for late-arriving facts.

A production team running dbt incremental models discovered that incremental predicates failures show up only when upstream dependencies shift traffic mix—staging load tests with uniform QPS missed the regression until Black Friday.

Runbook entry for incremental predicates: confirm blast radius (single namespace vs fleet-wide), identify last config change, roll back via documented single step, then capture metrics screenshots for postmortem—not ad-hoc dashboard hunting.

For dbt incremental models, instrument incremental predicates with low-cardinality metrics tied to user-visible outcomes: error rate, tail latency, freshness, or cost per successful operation—avoid paging on vanity gauges that never correlated with past incidents.

Game day scenario for incremental predicates: inject partial outage in staging quarterly, verify on- call can execute rollback in under fifteen minutes using only the linked runbook, update runbook with what actually broke.

Ownership for incremental predicates belongs in the service catalog with named rotation, last drill date, and known sharp edges—new engineers should deploy a safe canary within one week using that doc alone.

Change management for dbt incremental models: require peer review from someone outside the authoring team before production promotion—fresh eyes catch assumptions embedded in incremental predicates configs that authors no longer notice.

Capacity planning note: estimate peak QPS or job concurrency for incremental predicates, multiply by headroom factor one-point-five to two, compare against cloud quotas and license limits before launch week—not during the first outage.

Idempotent retries

unique_key required for merge—without it duplicates silently replay after failed runs.

A production team running dbt incremental models discovered that idempotent retries failures show up only when upstream dependencies shift traffic mix—staging load tests with uniform QPS missed the regression until Black Friday.

Runbook entry for idempotent retries: confirm blast radius (single namespace vs fleet-wide), identify last config change, roll back via documented single step, then capture metrics screenshots for postmortem—not ad-hoc dashboard hunting.

For dbt incremental models, instrument idempotent retries with low-cardinality metrics tied to user- visible outcomes: error rate, tail latency, freshness, or cost per successful operation—avoid paging on vanity gauges that never correlated with past incidents.

Game day scenario for idempotent retries: inject partial outage in staging quarterly, verify on-call can execute rollback in under fifteen minutes using only the linked runbook, update runbook with what actually broke.

Ownership for idempotent retries belongs in the service catalog with named rotation, last drill date, and known sharp edges—new engineers should deploy a safe canary within one week using that doc alone.

Change management for dbt incremental models: require peer review from someone outside the authoring team before production promotion—fresh eyes catch assumptions embedded in idempotent retries configs that authors no longer notice.

Capacity planning note: estimate peak QPS or job concurrency for idempotent retries, multiply by headroom factor one-point-five to two, compare against cloud quotas and license limits before launch week—not during the first outage.

Testing

Unit test SQL for is_incremental true/false; integration test retry produces identical row counts.

A production team running dbt incremental models discovered that testing failures show up only when upstream dependencies shift traffic mix—staging load tests with uniform QPS missed the regression until Black Friday.

Runbook entry for testing: confirm blast radius (single namespace vs fleet-wide), identify last config change, roll back via documented single step, then capture metrics screenshots for postmortem—not ad-hoc dashboard hunting.

For dbt incremental models, instrument testing with low-cardinality metrics tied to user-visible outcomes: error rate, tail latency, freshness, or cost per successful operation—avoid paging on vanity gauges that never correlated with past incidents.

Game day scenario for testing: inject partial outage in staging quarterly, verify on-call can execute rollback in under fifteen minutes using only the linked runbook, update runbook with what actually broke.

Ownership for testing belongs in the service catalog with named rotation, last drill date, and known sharp edges—new engineers should deploy a safe canary within one week using that doc alone.

Change management for dbt incremental models: require peer review from someone outside the authoring team before production promotion—fresh eyes catch assumptions embedded in testing configs that authors no longer notice.

Capacity planning note: estimate peak QPS or job concurrency for testing, multiply by headroom factor one-point-five to two, compare against cloud quotas and license limits before launch week—not during the first outage.

Operations

Alert on merge bytes processed ten times baseline—often missing predicate or accidental full refresh.

A production team running dbt incremental models discovered that operations failures show up only when upstream dependencies shift traffic mix—staging load tests with uniform QPS missed the regression until Black Friday.

Runbook entry for operations: confirm blast radius (single namespace vs fleet-wide), identify last config change, roll back via documented single step, then capture metrics screenshots for postmortem—not ad-hoc dashboard hunting.

For dbt incremental models, instrument operations with low-cardinality metrics tied to user-visible outcomes: error rate, tail latency, freshness, or cost per successful operation—avoid paging on vanity gauges that never correlated with past incidents.

Game day scenario for operations: inject partial outage in staging quarterly, verify on-call can execute rollback in under fifteen minutes using only the linked runbook, update runbook with what actually broke.

Ownership for operations belongs in the service catalog with named rotation, last drill date, and known sharp edges—new engineers should deploy a safe canary within one week using that doc alone.

Change management for dbt incremental models: require peer review from someone outside the authoring team before production promotion—fresh eyes catch assumptions embedded in operations configs that authors no longer notice.

Capacity planning note: estimate peak QPS or job concurrency for operations, multiply by headroom factor one-point-five to two, compare against cloud quotas and license limits before launch week—not during the first outage.

{% if is_incremental() %}
  where event_time >= (select max(event_time) - interval '3 days' from {{ this }})
{% endif %}

Choose merge with explicit unique_key for idempotent retries. Monitor merge bytes processed—ten× baseline often indicates missing incremental predicate or full-refresh accident.