Dbt Prod Ci Regression

Spellbook note: see also .cursor/skills/debug-ci/SKILL.md for CI context. Never embed API keys in the skill.

Run and validate (Dune MCP)

After SQL is built from branch context (PR merge SHA or correct github.sha source + {schema}_{alias}, filters):

1. Read MCP tool schemas first (required): under the workspace mcps/user-dune/tools/ (or the enabled Dune server’s tool descriptors), open the JSON for each tool you call so arguments match the contract.
2. createDuneQuery: create a temporary query (is_temp defaults true) with a clear name, the full DuneSQL query text, and optional description. Capture the returned query_id.
3. executeQueryById: run with that query_id (and performance if needed). Capture execution_id from the response.
4. getExecutionResults: pass executionId (ULID), increase timeout for heavy scans, use limit for previews. Interpret state: COMPLETED → check data.rows and resultMetadata.totalRowCount; FAILED → use errorMessage / errorMetadata to fix SQL and repeat.

Run exploratory SQL (min dates, raw counts) and the final diff query through the same pipeline. If MCP is unavailable, fall back to python scripts/dune_query.py (repo root, DUNE_API_KEY in .env).

Partitioning: Dune tooling expects filters on partition columns (e.g. block_date) where applicable—keep regression windows as tight as the comparison allows.

Investigate modified schema

1. Use git diff, dbt compile, _schema.yml, and model SQL to list relevant columns (partition keys, block_date, block_time, block_number, metrics like amount_usd, etc.). When in doubt, inspect the Dune table schema or compiled SQL for CI and prod.
2. Default comparison metrics (when columns exist): count(1) (rows) and sum(amount_usd) (or the spell’s primary USD column). That pair usually surfaces pipeline issues quickly. Extend or swap metrics from the schema (e.g. sum(amount_raw), volume fields) per lineage—tokens-style spells are the template, not a universal rule.
3. Choose a grain (group by / join keys) that exists on both CI and prod. It is not always blockchain: common Spellbook grains include blockchain, block_month, block_date, project, or combinations (e.g. blockchain + block_date). Match what the spell actually keys on for the question you are answering.
4. Coarse-first, then deeper: use high-level aggregates (single chain, or block_month / block_date slices) to confirm totals and distributions before joining on a row-level unique key (unique_key, tx_hash + evt_index, etc.). Cheap grain checks catch most pipeline regressions; unique-key diffs are for pinpointing survivors.
5. If there is no USD column, use count(1) and other numeric columns that make sense from the schema.

Query workflow (order matters)

Dynamic filters (required): Do not copy example block_date / block_number literals from docs or chat. Always run phase 1 on Dune (or MCP), read the result set, then substitute the returned bounds into every later query. Examples in reference.md use placeholders <MIN_BLOCK_DATE>, <MIN_BLOCK_NUMBER> for that reason.

1. Phase 1 — bounds on CI only (execute first, every run):
   select min(block_date) as min_block_date, min(block_number) as min_block_number from <ci_table> where block_date != current_date
   (Add extra predicates only if you need a narrower probe.) Use the result as shared lower bounds for both CI and prod in phases 2+.
2. Shared where fragment (same on all CTEs):
   block_date != current_date and block_date >= date '<MIN_BLOCK_DATE>' and block_number >= <MIN_BLOCK_NUMBER> — adjust or drop block_number if the model is not chain-scoped that way.
3. Parity check: count(1) and sum(<metric>) on prod and ci with the identical where; row count and metric totals should match before trusting a grain inner join.
4. Aggregate CTEs: same where, same group by grain, same metrics (e.g. count(1), sum(amount_usd)).
5. Join: inner join on the full grain (all group by columns). Use full outer join temporarily to inspect missing keys.
6. Diff filter last: abs(diff_rows) > 0 or abs(diff_usd) > <tolerance> (e.g. 5 for float noise) only after spot-checking unfiltered join output.
7. Validate: execute via Dune MCP; 0 rows in the diff-filtered grain query ⇒ pass at that tolerance.

Analysis outputs (what to report)

Build the shared where from phase 1 bounds (see workflow). Keep it on every CTE; join only on grain keys.

Optional: count(distinct …) per grain column on each side vs inner-join count.

See reference.md for phase 1 SQL, placeholder filters, grain diff query, and a rollup that returns counts + USD in one row.

Template (fill placeholders)

with ci as (
	select
		<grain_columns>
		, count(1) as total_rows
		, sum(<metric_column>) as total_metric
	from
		test_schema.git_dunesql_<GIT_HASH>_<schema>_<alias>
	where
		block_date != current_date
		and block_date >= date '<MIN_BLOCK_DATE>'
		and <optional_aligned_predicates_e_g_block_number>
	group by
		<grain_columns>
)
, prod as (
	select
		<grain_columns>
		, count(1) as total_rows
		, sum(<metric_column>) as total_metric
	from
		<prod_catalog>.<prod_schema>.<prod_alias>
	where
		block_date != current_date
		and block_date >= date '<MIN_BLOCK_DATE>'
		and <same_optional_predicates_as_ci>
	group by
		<grain_columns>
)
select
	<compare_columns>
from prod
inner join ci
	on <join_keys>
where
	(
		abs(prod.total_rows - ci.total_rows) > 0
		or abs(prod.total_metric - ci.total_metric) > <tolerance>
	)
order by
	1

Cross-repo use

The same workflow applies in any dbt + Dune (or Trino) repo: adjust prod relation (catalog.schema.table), CI schema/table prefix, and column names. Copy this folder to ~/.cursor/skills/dbt-prod-ci-regression/ for a personal default.

More examples

See reference.md for a filled tokens.transfers example and exploratory snippets.