All committers have signed the CLA.

Workflow [PR], commit [85a2f29]

Glad you're already working on this @htuall 🎉! Please next time ping me directly or even assign the PR to me so that I'm aware. Somehow this went off under my radar till now 👀

Have you checked how the llvm-xray tool is doing this same thing? I assumed we could get the mapping the same way and our part would be to find the proper function addresses based on non-mangled function names, which is what I think readSymbolTable is doing, right? We'll probably need to demangle those, or alternatively mangle the input symbols to be able to match them. The latter seems less work, since you only need to do that for those symbols we'll be interested in, rather than converting all symbols.

Also, I'm afraid I can't yet understand Russian comments that well even after @nikitamikhaylov's efforts to teach me the basics. So, please use English whenever possible so that I don't have to be going back and forth translating 🙏

Creating a Chrome trace format from systen.xray_instrumentation_profiling_log:

Ask to profile and sleep a bit for every query entry to have some data:

SYSTEM INSTRUMENT ADD `QueryMetricLog::startQuery` PROFILE
SYSTEM INSTRUMENT ADD `QueryMetricLog::startQuery` SLEEP ENTRY 0.5

Send 1000 parallel queries using 24 parallel jobs:

yes "clickhouse client --query 'select 1'" | head -n 1000 | parallel -j24

The chrome_trace.sql query:

SELECT
    format(
        '{{"traceEvents": [{}]}}',
        arrayStringConcat(
            groupArray(
                concat(
                    -- Begin Event
                    format(
                        '\n{{"name": "{}", "cat": "{}", "ph": "B", "ts": {}, "pid": 1, "tid": {}, "args": {{"query_id": "{}"}}}}',
                        name,
                        'clickhouse',
                        toString(event_time_microseconds),
                        toString(tid),
                        query_id
                    ),
                    ',',
                    -- End Event
                    format(
                        '\n{{"name": "{}", "cat": "{}", "ph": "E", "ts": {}, "pid": 1, "tid": {}, "args": {{"query_id": "{}"}}}}',
                        name,
                        'clickhouse',
                        toString(event_time_microseconds + duration_microseconds),
                        toString(tid),
                        query_id
                    )
                )
            ),
            ','
        )
    ) AS trace_json
FROM system.xray_instrumentation_profiling_log
WHERE event_date >= today();

clickhouse client --query "$(cat chrome_trace.sql)" > trace.json

trace.json

We can see in Perfetto

Comparing the binary size:

Random PR:

• AMD release: 153240109 bytes
• ARM release: 144269892 bytes

This PR:

• AMD release: 160552808 bytes
• ARM release: 154562732 bytes

For x86_64, there's a ~4.5% of binary size increase. For aarch64, there's a ~7% binary size increase.

It seems we cannot use XRay and any other sanitizer at the same time as per https://s3.amazonaws.com/clickhouse-test-reports/json.html?PR=77153&sha=23c6ed27502189ca2df435a42c653cb5ed18896c&name_0=PR

Even though I haven't found on the internet any evidence of them being incompatible, both sanitizers and XRay seem to use the same underlying allocator and runtime infrastructure from sanitizer_common, which is why you get "multiple definition" errors for symbols like __sanitizer::internal_allocator().

I'm closing this PR in favor of a new one where I'm the owner and I can freely change the description, changelog and everything without impersonating @htuall. Having to approve the workflow for every commit was also bad.

I'm keeping all the commits from Alina and will provide credit to her 💪

#89173