Record metrics and charge for HTTP route egress#5505
Conversation
| // Charge egress, using BSATN size for parity with WebSocket queries. | ||
| metrics.bytes_sent_to_clients += rows.iter().map(|row| bsatn::to_len(row).unwrap_or(0)).sum::<usize>(); |
There was a problem hiding this comment.
This seems potentially redundant in a way that worries me. If we charge for both bytes_sent_to_clients and http_response_body_bytes, then customers get charged for egress due to SQL queries over HTTP twice. If we call run_inner from the WS path (which I don't think we do, at least currently, but I could easily imagine us doing in the future), then we'll double count on bytes_sent_to_clients directly.
There was a problem hiding this comment.
Good to know. I will look into better future-proofing this, if that works.
There was a problem hiding this comment.
@gefjon Thanks reviewing + catching that. Moved it into exec_sql, so run_inner doesn't touch egress anymore and each transport charges for what it sends (like WS already does...). Also added a test to make sure the executor never charges. The http_* metrics aren't billed, they're just for observability. Added a comment for clarity.
Each transport charges for the rows it sends, matching how WebSocket queries charge at their serialization point. Prevents double counting if the executor gains other transport callers.
2 participants
Description of Changes
The HTTP
/sqlendpoint charged for its DB work (rows/bytes scanned, writes) but not for egress, the bytes sent back to the client. WebSocket queries already count these, but the HTTP path skipped it, so the same query was billed over WS and free over HTTP. There were also no metrics on the HTTP routes.This PR:
/sql, using the BSATN size of the result rows so the price matches WebSocket queries./v1middleware that records per-route request count, latency, and body sizes. Body sizes count bytes actually transferred through a smallBodywrapper rather than trusting the Content-Length header, so chunked and streamed bodies are counted and spoofed headers are not. Non-standard HTTP methods are bucketed into anOTHERlabel to keep cardinality bounded.Adds
http-bodyas a direct dependency (already in the tree via axum/hyper).Not covered here: CORS preflights and unmatched 404s are not counted (they short-circuit before the middleware), and routes other than
/sqlare not billed egress yet.API and ABI breaking changes
None. The
/sqlresponse is unchanged, only internal metrics were added.Expected complexity level and risk
Testing
SELECTchargesbytes_sent_to_clientsequal to the rows' BSATN size.SELECTover HTTP bumps the egress counter by the exact BSATN size and thespacetime_http_*counters by the exact wire bytes (checked against curl's size_download).curl -X BANANA) lands in theOTHERbucket, and a chunked request body (no Content-Length) is counted.