sql: add request unit to tracing sampled query event

[j82w]

Fixes: #95194

Release note (sql change): Add the request unit estimate to the sampled query event.

[cockroach-teamcity]

This change is 

[DrewKimball]

I don't think we can do this yet, since the network contribution to RU consumption is only measured for EXPLAIN ANALYZE right now, not all sampled queries. Also worth mentioning that the estimate is currently only shown for tenants, and only for vectorized execution.

[kevin-v-ngo]

Hi @DrewKimball, do you know how we get network information (bytes) today for SQL stats? Is this a different network metric used for RUs today?

Just want to understand relationship so we have consistent/correlated metrics

[DrewKimball]

Hi @DrewKimball, do you know how we get network information (bytes) today for SQL stats? Is this a different network metric used for RUs today?

Do you mean NetworkBytesSent? I believe that measures the number of bytes sent internally during execution of a SQL query from node to node, not from node to client. It also doesn't include remote requests made through KV (e.g. for a scan doing some remote reads). RUs currently don't care about internal network usage, only the bytes sent to the client.

During normal execution we track the bytes sent to the client during normal execution as the results are encoded. During a command like EXPLAIN ANALYZE, the results aren't actually sent to the client, so instead we pass them through a helper that encodes the result, measures the size, and then discards the encoded bytes (see here). Since estimated RUs have so far only been set up to work with EXPLAIN ANALYZE, they aren't hooked up to the network egress measured during normal execution.

[kevin-v-ngo]

Hi @DrewKimball, do you know how we get network information (bytes) today for SQL stats? Is this a different network metric used for RUs today?

Do you mean NetworkBytesSent? I believe that measures the number of bytes sent internally during execution of a SQL query from node to node, not from node to client. It also doesn't include remote requests made through KV (e.g. for a scan doing some remote reads). RUs currently don't care about internal network usage, only the bytes sent to the client.

During normal execution we track the bytes sent to the client during normal execution as the results are encoded. During a command like EXPLAIN ANALYZE, the results aren't actually sent to the client, so instead we pass them through a helper that encodes the result, measures the size, and then discards the encoded bytes (see here). Since estimated RUs have so far only been set up to work with EXPLAIN ANALYZE, they aren't hooked up to the network egress measured during normal execution.

Thank you @DrewKimball for the explanation. I've filed a related issue here tracking better observability into RUs: #108436.

[j82w]

Closing this based on the feedback on ru accuracy.