dbts

dbt environment runner with Snowflake zero-copy clone sandboxes.

Run dbt against a private, per-developer Snowflake clone of staging or live without managing a separate profile or CLI — and use the same tool as a single front for your shared dev, staging, and live targets.

Install

uv tool install dbts
# or run ad hoc:
uvx dbts ...

Quick start

1. Add a sandbox: target to ~/.dbt/profiles.yml (alongside your existing dev, staging, live). The database name must match the pattern <PREFIX>_SANDBOX_<USER>:

   <your_profile_name>:           # whatever your dbt_project.yml's profile: field says
     outputs:
       sandbox:
         type: snowflake
         account: <same as the other targets>
         user: <same as the other targets>
         role: <same as the other targets>
         authenticator: externalbrowser
         database: <prefix>_sandbox_<your_username>
         warehouse: <same as the other targets>
         schema: <same as the other targets>
   

2. Create your clone:

   dbts up --from staging
   

3. Run dbt against it:

   dbts build my_model
   dbts test  +my_model+
   

4. Refresh or drop when done:

   dbts refresh --from staging
   dbts drop
   

Commands

Sandbox — manage the per-developer zero-copy clone:

dbts up        --from staging|live         create the clone
dbts refresh   --from staging|live         drop and re-create the clone
dbts status                                show clone DB, source, age, owner
dbts drop                                  drop the clone
dbts publish                               grant ROLE PUBLIC read access (incl. future objects)
dbts unpublish                             revoke ROLE PUBLIC's access

Inspect — preview / audit before (or after) a build:

dbts plan      [selectors...]              preview the build set
  --cost                                     estimate credits + runtime from QUERY_HISTORY
  --days N                                   QUERY_HISTORY lookback window (default 7, max 365)
  --output console|html                      render output (html writes an interactive DAG)
  --output-file PATH                         html output path (default ./plan-graph.html)
dbts freshness [selectors...]              audit lineage; flag stale tables, suggest a rebuild
  --since                                    explicit threshold (24h, 7d, 1w, or ISO datetime)
  --output console|html                      render output (html writes an interactive DAG)
  --output-file PATH                         html output path (default ./freshness-graph.html)

dbt pass-through — forwarded to dbt with --target sandbox by default:

dbts run|build|test|compile|seed|snapshot|ls|show
dbts debug|deps|source|docs|parse|clean
  --target sandbox|staging|live|dev          choose target (-t)

Meta:

dbts version                               print installed version

Global flags: -v / --verbose (debug logging, including DDL), -q / --quiet (warnings only). -h is a shorthand for --help on every command. Run dbts <command> --help for the full option list.

Selectors

Bare positional model selectors work the same as in dbt:

dbts build my_model+              # forwarded as `--select my_model+`
dbts test +my_model+              # ancestors and descendants
dbts run a b c+                   # multi-selector union
dbts build my_model+ --exclude experiments
dbts build --select a b           # `--select` + bare positional are merged

Bare positional args on run / build / test / compile / seed / snapshot / ls / show are promoted to --select before being forwarded. Other subcommands (debug, deps, docs, parse, clean, source) pass arguments through verbatim.

Previewing a build

dbts plan lists exactly which models a dbts build (or run/test/...) with the same selectors would touch. Useful when a build against --target live blows up halfway and you need to add --exclude rules.

dbts plan my_model+ another_model+ --target live           # offline, fast
dbts plan --select tag:slow --exclude path:models/intermediate
dbts plan my_model+ --target live --cost                   # + Snowflake cost breakdown

By default the command is offline — it parses the dbt project but never connects to Snowflake. Output groups models by directory and shows materialization, tags, and parent count. The footer prints suggested --exclude path:<dir> and --exclude tag:<name> snippets sized by how many models each would prune.

Pass --cost to also estimate Snowflake credits and runtime from SNOWFLAKE.ACCOUNT_USAGE.QUERY_HISTORY. Each row gets median run and last seen columns; the footer adds total credits + USD for an incremental vs full refresh, plus a top-5 most expensive list. Default lookback is 7 days (--days N, max 365).

Cost estimates require a structured query_tag with a model field (HelloFresh's set_query_tag macro is one example) and read access to SNOWFLAKE.ACCOUNT_USAGE.QUERY_HISTORY. USD uses $3.00/credit by default; override with $DBTS_CREDIT_RATE. If access is missing, dbts plan falls back to the offline output.

Graph view (--output html)

dbts plan tag:hourly --output html               # writes ./plan-graph.html (topology only)
dbts plan tag:hourly --cost --output html        # nodes coloured by p50 runtime / credits
dbts plan tag:hourly --output html --output-file /tmp/plan.html

Renders the build set as an interactive top-down DAG (vis-network, loaded from unpkg.com). Nodes scale with median runtime and shade red as runtime climbs; pass --cost for cost-based colouring (offline runs render neutral grey). Hover for a tooltip with materialization, parents, p50 runtime, estimated credits/USD, and tags. Type in the search box to filter by name. Existing files are overwritten silently.

Auditing freshness after an incident

dbts freshness answers "did data flow correctly through this lineage?" by reading INFORMATION_SCHEMA.TABLES.LAST_ALTERED for every table in the selected build set, sorting topologically (parents before children), and highlighting stale links in red.

# Did the catch-up run reach everything downstream of these models? (defaults to --target live)
dbts freshness base__recipe__cps+ base__recipe__ingredient__cps+

# Audit the full lineage in both directions
dbts freshness +base__recipe__cps+

# Compare against an explicit incident timestamp
dbts freshness base__recipe__cps+ --since '2026-05-09 17:00'

# Audit your sandbox clone instead
dbts freshness base__recipe__cps+ --target sandbox

Default target: live (unlike build / plan, which default to sandbox) — the typical post-incident question is "did production catch up?". Pass --target sandbox|staging|dev to audit elsewhere. The command resolves the target to the physical database name (<DB>_SANDBOX_<USER>, <DB>_STAGING, <DB>_DEV, or unsuffixed for live), mirroring the project's generate_database_name macro.

How "stale" is decided

Two values drive the colored output:

• Baseline — the most recent LAST_ALTERED across all tables in the build set. It's the "freshest thing you have right now," shown in the header for context.
• Threshold — anything older than this is flagged as stale (red row).

Without --since, the threshold is adaptive: baseline − 6 hours. The threshold drifts with whatever's currently fresh, so you don't have to know what "fresh" means today.

dbts freshness recipe+
# baseline = 14:02 (freshest table)
# threshold = 08:02   (anything older is stale)

With --since, the threshold is explicit and the adaptive window is ignored. Useful when you know exactly when something broke:

dbts freshness recipe+ --since '2026-05-09 17:00'   # absolute (ISO datetime or date)
dbts freshness recipe+ --since 24h                  # relative (s, m, h, d, w)

The footer prints a copy-pasteable dbts build --select X+ line that covers the minimum set of stale-roots needed to catch the chain back up — building any model that's already fresh upstream is avoided.

Graph view (--output html)

dbts freshness recipe+ --output html             # writes ./freshness-graph.html
dbts freshness --since 24h --output html --output-file /tmp/freshness.html

Renders the lineage as a top-down DAG with stale tables in red, fresh in green, missing/no-data in grey. Edges between two stale nodes are coloured red so a broken chain is visible at a glance. Hover for database.schema, last_altered, time-ago and row count; type in the search box to filter.

What LAST_ALTERED actually means

The signal is Snowflake's INFORMATION_SCHEMA.TABLES.LAST_ALTERED, which dbt bumps on every INSERT, MERGE, UPDATE, or CREATE OR REPLACE. "Fresh" means "dbt touched it recently," not "new rows arrived" — exactly the right semantic for the "did the chain re-execute?" question. Requires read access to INFORMATION_SCHEMA.TABLES.

Project-side coupling

dbts assumes the dbt project's generate_database_name macro recognises ENV=sandbox and routes models into a _SANDBOX_<USER> suffixed database. See the dbt project's README for the macro snippet.

Profile resolution

dbts resolves the dbt profile name in this order:

1. $DBTS_PROFILE if set.
2. The profile: field in dbt_project.yml at the project root.

Jinja {{ env_var('NAME', 'default') }} calls in the profile: field are rendered against the current environment, so projects whose profile name is templated (e.g. tardis_{{ env_var('warehouse', 'snowflake') }}) work out of the box.

Development

uv sync --group dev
uv run pytest
prek install   # one-time, runs ruff + ty on every commit

prek (Astral's Rust port of pre-commit) runs the hooks in .pre-commit-config.yaml. Install via brew install prek or uv tool install prek; the standard pre-commit binary also works.

Cutting a release (after moving [Unreleased] entries under a [X.Y.Z] heading in CHANGELOG.md):

./scripts/release.sh 0.6.0

The script bumps pyproject.toml, syncs the lockfile, runs checks, commits, tags, pushes, and creates the GitHub release from CHANGELOG.md. See scripts/README.md for the full workflow and recovery tips.