ducklens · warehouse fit report

ducklens

Warehouse Fit Report · snowflake
2026-03-17 to 2026-06-14 (90d) · 25,000 queries · 5 warehouses

What fits on one machine

39%of your Snowflake bill is query compute that fits on one DuckDB machine

$43,442/yr movable of a $112,348/yr invoice · 69% of attributed query compute · 43% of queries

Recommendation2 MOVE2 SPLIT1 KEEPof 5 warehouses

$112,348

bill analyzed / yr

$43,442

movable to DuckDB / yr

$9,652

runs at (box + storage) / yr

Where your Snowflake bill goes

$112,348/yr, every dollar colored by what to do with it. A cost dashboard stops at the first column.

Movable to DuckDB$43,442/yr · 39%

Must stay (residual)$19,298/yr · 17%

Idle warehouse time$37,286/yr · 33%

Serverless / cloud services$12,322/yr · 11%

Gross annual saving

$28,722 - $38,859 / yr

central $33,791 run-rate delta: migrate $43,442/yr of compute, run it at ~$9,652/yr. Excludes one-time migration, dual-running, egress, at-rest storage, ops.

Hybrid split

Each warehouse scored on its own bill: MOVE wholesale, SPLIT the movable queries off, or KEEP as is.

Warehouse	Fit	Spend/yr	Verdict	Top blocker
AD_HOC_WH	77%	$18,665	SPLIT	warehouse-specific SQL (heuristic)
ANALYTICS_WH	79%	$17,103	SPLIT	warehouse-specific SQL (heuristic)
BI_SERVING_WH	37%	$15,032	KEEP	high-concurrency serving
DBT_WH	85%	$8,926	MOVE	multi-statement write transactions
LOAD_WH	81%	$3,015	MOVE	continuous high-frequency writes

Before a MOVE, clear: inbound/outbound Secure Data Shares · RBAC/row-access/masking policies · downstream BI & reverse-ETL · Streams/Tasks/Dynamic Tables · data residency/compliance.

What must stay, and why

Reason	Spend/yr	Queries
high-concurrency serving	$5,902	6,048
warehouse-specific SQL (heuristic)	$5,784	899
multi-statement write transactions	$3,295	687
multi-cluster scale-out	$1,469	798
stored procedures / UDFs	$1,317	177
continuous high-frequency writes	$1,045	5,312
queued / overloaded	$485	273

Top blockers

The costliest queries keeping a warehouse on Snowflake.

Query	Warehouse	Reason	Spend/yr
982355990f72... CREATE TABLE stg_orders AS SELECT * FROM raw.orders	AD_HOC_WH	multi-statement write transactions	$57
01e3f6d154b7... INSERT INTO target SELECT * FROM staging	AD_HOC_WH	multi-statement write transactions	$44
794ec926bc9e... INSERT INTO target SELECT * FROM staging	AD_HOC_WH	multi-statement write transactions	$42
be806943707d... SELECT * FROM TABLE(GENERATOR(ROWCOUNT => 1000))	AD_HOC_WH	warehouse-specific SQL (heuristic)	$41
ed22f249555b... INSERT INTO target SELECT * FROM staging	AD_HOC_WH	multi-statement write transactions	$41
97801d0e3e1f... INSERT INTO target SELECT * FROM staging	AD_HOC_WH	multi-statement write transactions	$41
280f47eb33bc... CALL refresh_marts()	AD_HOC_WH	stored procedures / UDFs	$40
e3803de21a67... SELECT PARSE_JSON(payload):user:id::string AS uid FROM events	AD_HOC_WH	warehouse-specific SQL (heuristic)	$40

If the box had more disk

Fit recomputed at other local-NVMe sizes; DuckDB spills to disk, so disk is the spill budget.

Local NVMe	Fit % of query compute	Movable $/yr
2 TB	69%	$43,419
4 TB	69%	$43,442
7.5 TB (target)	69%	$43,442
8 TB	69%	$43,442
16 TB	69%	$43,442
30 TB	69%	$43,442

Assumptions

256 GB RAM + NVMe box at $800/mo (an r6id.4xlarge-class instance, reserved, incl. ops headroom) + object-storage reads at $0.50/TiB, full re-read assumed (no cache credit). Pricing basis: $3.00/credit (enterprise). The ±15% band is illustrative, not a confidence interval.

Caveats

Totals reconcile to METERING_DAILY_HISTORY; the non-warehouse (serverless/cloud-services) bucket is shown separately and is NOT migrated by this analysis.
Per-query cost is an attribution of the metered total (it sums back to it), not a marginal cost; movable spend only materializes when a warehouse can be downsized or suspended.
SQL-dialect portability is assessed heuristically; Snowflake-specific functions may still block a query the structural flags pass. The org-blocker checklist must be cleared per MOVE.
DuckDB spills to local disk (out-of-core), so 'too big for RAM' is not 'too big for one machine'; a query fits unless its spill exceeds the target box's local NVMe, and disk-spilling is slower than in-memory. Fit is led by observed spill, not raw scan size.
The annual figure is period x (365/days), NOT seasonally adjusted, so month/quarter-end peaks may not be represented; review at least one full month-end before trusting it.
The saving is a GROSS run-rate delta: it excludes one-time migration cost, dual-running during cutover, egress, at-rest storage, and the operational surface you take on.
Account Usage latency is up to ~3h, so the last few hours may be incomplete.