feat: finops_snowflake_savings: read-only Snowflake warehouse savings check
Summary (for people)
- Why: this does for Snowflake what #1762 does for Databricks. A user can check their own Snowflake account for warehouse savings from their machine, or from Claude Code's
/savings, without sending usage data to Altimate. - Before ā after: Before, altimate-code had no Snowflake savings check. After,
finops_snowflake_savingsreads seven days of the account's usage. It returns each saving with a monthly dollar range and its evidence. The checks cover warehouses that sit idle, could run one size smaller, keep an unneeded extra cluster, or restart so often that a billed minute is paid twice. - Who is affected: Users with a Snowflake connection, once this ships. The check runs only with
confirm: true, and the agent asks first, because a run can resume a suspended warehouse that then bills. Other Snowflake queries are unchanged: the new cancel path is opt-in. - Before it can merge:
Details (for agents/reviewers)
Issue for this PR
None.
Type of change
- Bug fix
- New feature
- Refactor / code improvement
- Documentation
What does this PR do?
- Facts. The tool lists warehouses with
SHOW WAREHOUSES. To report on seven days, it reads up to eleven days ofSNOWFLAKE.ACCOUNT_USAGE. Days are UTC, and the window ends at least six hours ago, past the views' latency. - Rules. altimate-core's
snowflakeSavingsdecides what qualifies and what it saves. It returns the same report shape as the Databricks tool. - Price. A credit is priced from the account's rate sheet when it can be read in dollars, otherwise at a stated $3.00.
- Asks first. Without
confirm: truenothing runs. With it, the check stops after 9 minutes, connecting included, or sooner at the run's time budget. - Stopping for real. The Snowflake driver gains the opt-in
cancelAfterMs. Past the limit, or when the session stops, the statement is cancelled on Snowflake. A statement still waiting on a reconnect is never sent. - Assessment mode. When it loads, snowflake-sdk probes the AWS, Azure and GCP metadata services. In this mode the driver sets
SNOWFLAKE_DISABLE_PLATFORM_DETECTIONfirst, so the SDK skips them. - Limits, as the hosted product has them. An active period starts at a resume, ends at the next event, and must overlap at least one query, as the core rules' spec defines it. So a warehouse that never suspended in the window, or a period with no query, gets no idle finding. A bounce counts within one UTC day.
- Wiring and docs. The tool is registered where
finops_databricks_savingsis. Docs are on the FinOps tools page.
How did you verify your code works?
At head ade5a6b5e6, rebased onto #1762's current head:
tsgo --noEmit:packages/opencodeandpackages/driversboth clean.- Drivers,
databricks-deadlineandsnowflake-deadline: 26 pass. - The Snowflake and Databricks savings suites, plus the analyst, workspace and telemetry suites: 1,853 pass in 61 files.
assessment-zero-egress: 23 pass. That includes the Snowflake case with the real driver loaded: without assessment mode the proxy sees the metadata probes, and with it only the account.- Live, on Altimate's own Snowflake account, with the packaged CLI, a throwaway home and a read-only role: 9 findings in 43 s. The only hosts contacted were the Snowflake account and the Databricks workspace. The call without
confirmcontacted nothing. - Against stored data. For a past week, the measured credits equal the hosted product's on all nine active warehouses, and all 1,275 stored active periods match.
- Certificate revocation. The SDK's revocation requests use an HTTPS agent that Bun does not use, so the packaged CLI makes none.
- Large results. A result big enough for Snowflake to serve from its cloud storage would add that host. The live run did not need it, and the docs say so.
- Not covered: cancellation at the limit was not exercised live. The driver tests cover it with a fake SDK. The live run is repeated on the final head before merge.
Cost side effects
Yes, warehouse queries. A confirmed run sends six statements:
SHOW WAREHOUSES, which needs no warehouse;- five queries on the connection's warehouse, which Snowflake resumes if it is suspended and auto-resume is on.
The usage queries scan up to eleven days of ACCOUNT_USAGE, so their work grows with query history. They are capped by the explicit confirm, one 9-minute limit for the whole run with server-side cancel, and a row limit per query. No model calls.
Screenshots / recordings
Not a UI change.
Checklist
- I have tested my changes locally
- I have not included unrelated changes in this PR
- I checked for cost side effects (model calls, queries or cloud resources that grow with user data) and described them, or wrote "none"
š¤ Generated with Claude Code
Summary by cubic
Adds the finops_snowflake_savings check, the Snowflake counterpart to the Databricks savings tool. Users can check their own account for warehouse savings from their machine or the /savings tool, without sending usage data to Altimate.
How it works
- Reads seven days of usage from
SHOW WAREHOUSESandSNOWFLAKE.ACCOUNT_USAGE, returning each saving with a monthly dollar range and evidence. - Catches idle warehouses, warehouses that could run one size smaller, unneeded extra clusters, and frequent restarts that double-bill a minute.
- Runs only with
confirm: trueand stops after 9 minutes, because a run can resume a suspended warehouse that then bills. - Prices a credit from the account's rate sheet when available, otherwise at $3.00.
- The Databricks savings check now also reports
workspaces_checked, counting distinct workspaces across every checked row, with or without a finding or SQL warehouse.
Driver changes
- The Snowflake driver gains opt-in
cancelAfterMs, cancelling statements server-side past the limit or when the run stops. - In assessment mode, the driver sets
SNOWFLAKE_DISABLE_PLATFORM_DETECTIONso the SDK skips its cloud metadata probes.
Depends on the Databricks savings PR (#1762) merging first, plus an altimate-core release that exports snowflakeSavings and a pin bump.