sql: incomplete vectorized traces

[glennfawcett]

Describe the problem

I ran across an interesting observation while trying to explain how queries are run in parallel within CRDB. I was using the internal KV workload and have a fair number of ranges distributed across three nodes:

root@:26257/kv> SELECT k2, count(*) FROM kv GROUP BY k2;
    k2   |  count
---------+----------
  z      | 4561620
  needle |       8
  

I was expecting that the above query would run in parallel, but when looking at the Jaeger trace collected, it appears to not be serial.

[stmt-bundle-625911811735650305.zip](https://github.com/cockroachdb/cockroach/files/5889843/stmt-bundle-625911811735650305.zip) [stmt-bundle-628463749489590273.zip](https://github.com/cockroachdb/cockroach/files/5889847/stmt-bundle-628463749489590273.zip)

At the advise of engineering, I ran this by disabling vectorization SET vectorize = off and it appeared to run as expected albeit longer.

[RaduBerinde]

To clarify - the query runs in parallel on the three nodes, but the traces for the actual flow and scans are missing from node 2 and 3. It looks like a problem with tracing when using the vectorized engine. I tried to repro locally in a simpler case and could not - I saw traces for scans from all nodes.

[yuzefovich]

Yeah, it is a known issue (#55821) that we intend to address for 21.1 release.

[yuzefovich]

An update here is that #61380 will fix it and will be in 21.1: