Request coalescing

When a dashboard refreshes, ten users open the same report at 9am, or a CI matrix fans out identical dbt queries, Snowflake sees N copies of the same query — each one a candidate to resume a warehouse and burn credits. chukei collapses them into one execution and fans the single result back out.

One execution, many waiters

  time --->

  Client A  --Q(F)-->+
                     │  chukei runs Q once  ->  Snowflake
  Client B  --Q(F)-->+  (B and C join the in-flight request,
                     │   so the warehouse runs Q just once)
  Client C  --Q(F)-->+
                     │
                     +  <------- single result for Q
                     │
                     +  --> result to A
                     +  --> result to B   (no extra cost)
                     +  --> result to C   (no extra cost)

Coalescing keys on the hard fingerprint: while a query with fingerprint F is in flight to Snowflake, any further request with the same F attaches to the pending execution instead of issuing its own. When the upstream result returns, every waiter is served from it.

Coalescing vs caching

They are complementary, and both subject to the determinism gate:

	Caching	Coalescing
Targets	repeated queries over time	identical queries at the same instant
Saves	the whole execution	the duplicate concurrent executions
Window	until invalidation	the in-flight duration only
Needs a prior result?	yes	no — the first one is still running

Coalescing captures the savings that caching can't: the first time a herd of identical queries arrives, there is no cache entry yet, but there is still only one real query worth running. This is a major lever for BI dashboard costs, where many viewers hit the same tiles simultaneously.

Safety

Only determinism-gate-clean reads are coalesced — writes and non-deterministic queries each execute independently, and if anything is uncertain the requests fail open and run separately. Coalescing never changes a result; it only removes redundant executions of an identical one.

Related

• Snowflake query caching guide
• Query fingerprinting — the key coalescing uses
• Architecture overview