Moving LogRecord implementation to SDK

[lalitb]

Fixes #1189

Changes

Opening as draft to get the initial feedback. The changes are:

• Created LogRecord traits in the API.
• Added corresponding struct LogRecord with implementation in the SDK.

Earlier:

        logger.emit(
            LogRecord::builder()
                .with_body("full log")
                .with_timestamp(now)
                .with_observed_timestamp(now)
                .with_severity_number(Severity::Warn)
                .with_severity_text(Severity::Warn.name())
                .with_attribute("name", "my-event-name")
                .with_attribute("event.id", 20)
                .with_attribute("user.name", "otel")
                .with_attribute("user.email", "otel@opentelemetry.io")
                .with_attribute("code.filename", "log.rs")
                .with_attribute("code.filepath", "opentelemetry_sdk/benches/log.rs")
                .with_attribute("code.lineno", 96)
                .with_attribute("code.namespace", "opentelemetry_sdk::benches::log")
                .with_attribute("log.target", "opentelemetry_sdk::benches::log")
                .build(),
        )

With PR:

        let mut log_record = logger.create_log_record();
        log_record.set_body("full log".into());
        log_record.set_timestamp(now);
        log_record.set_observed_timestamp(now);
        log_record.set_severity_number(Severity::Warn);
        log_record.set_severity_text(Severity::Warn.name().into());
        log_record.add_attribute("name", "my-event-name");
        log_record.add_attribute("event.id", 20);
        log_record.add_attribute("user.name", "otel");
        log_record.add_attribute("user.email", "otel@opentelemetry.io");
        log_record.add_attribute("code.filename", "log.rs");
        log_record.add_attribute("code.filepath", "opentelemetry_sdk/benches/log.rs");
        log_record.add_attribute("code.lineno", 96);
        log_record.add_attribute("code.namespace", "opentelemetry_sdk::benches::log");
        log_record.add_attribute("log.target", "opentelemetry_sdk::benches::log");
        logger.emit(log_record);

Note - with the new design, the log record creation through chaining required cloning eventually while emitting, so haven't implemented the chaining. The tracing appender also doesn't require chaining as all these log data are populated separately.

This allows us to change the internal SDK implementation for LogRecord without breaking the Bridge API. While there are no current perf benefits with the approach, it enables us to add more optimizations in the SDK to optionally allow the serialization of the log records in the export format without any intermediate allocations within the SDK. I plan to add those optimizations in subsequent PRs once this is finalized.

Benchmark

Test Name	Main	PR
simple-log/no-context	[105.28 ns]	[89.896 ns]
simple-log/with-context time	[115.49 ns]	[90.293 ns]
simple-log-with-int/no-context	[141.40 ns]	[123.36 ns]
simple-log-with-int/with-context	[147.15 ns]	[123.61 ns]
simple-log-with-double/no-context	[145.43 ns]	[123.32 ns]
simple-log-with-double/with-context	[153.33 ns]	[122.28 ns]
simple-log-with-string/no-context	[144.01 ns]	[122.61 ns]
simple-log-with-string/with-context	[149.85 ns]	[123.61 ns]
simple-log-with-bool/no-context	[146.57 ns]	[121.77 ns]
simple-log-with-bool/with-context	[152.56 ns]	[123.12 ns]
simple-log-with-bytes/no-context	[161.11 ns]	[139.43 ns]
simple-log-with-bytes/with-context	[171.69 ns]	[140.73 ns]
simple-log-with-a-lot-of-bytes/no-context	[164.77 ns]	[140.72 ns]
simple-log-with-a-lot-of-bytes/with-context	[172.50 ns]	[142.45 ns]
simple-log-with-vec-any-value/no-context	[205.88 ns]	[194.93 ns]
simple-log-with-vec-any-value/with-context	[214.82 ns]	[194.06 ns]
simple-log-with-inner-vec-any-value/no-context	[275.55 ns]	[256.61 ns]
simple-log-with-inner-vec-any-value/with-context	[280.79 ns]	[255.82 ns]
simple-log-with-map-any-value/no-context	[234.09 ns]	[207.90 ns]
simple-log-with-map-any-value/with-context	[244.29 ns]	[207.57 ns]
simple-log-with-inner-map-any-value/no-context	[335.55 ns]	[303.23 ns]
simple-log-with-inner-map-any-value/with-context	[344.18 ns]	[304.65 ns]
long-log/no-context	[106.17 ns]	[89.779 ns]
long-log/with-context	[118.93 ns]	[91.673 ns]
full-log/no-context	[143.31 ns]	[95.529 ns]
full-log/with-context	[145.20 ns]	[96.720 ns]
full-log-with-4-attributes/no-context	[315.31 ns]	[227.49 ns]
full-log-with-4-attributes/with-context	[312.81 ns]	[235.11 ns]
full-log-with-9-attributes/no-context	[548.88 ns]	[406.99 ns]
full-log-with-9-attributes/with-context	[501.88 ns]	[418.26 ns]
full-log-with-attributes/no-context	[322.71 ns]	[277.63 ns]
full-log-with-attributes/with-context	[328.19 ns]	[279.08 ns]

Stress Test (20 threads)
Main: ~35 million iterations/sec
PR: ~39 million iterations/sec

Merge requirement checklist

• CONTRIBUTING guidelines followed
• Unit tests added/updated (if applicable)
• Appropriate CHANGELOG.md files updated for non-trivial, user-facing changes
• Changes in public API reviewed (if applicable)

[TommyCpp]

Do we really need a builder for LogRecord? Looking at the trait all it provides are method to change fields. So LogRecord by itself looks like a builder to me.

[lalitb]

Yes, I think I over-engineered a bit. Have removed the builder trait now :)

[cijothomas]

While there are no current perf benefits with the approach,

Can you post benchmarks/stress tests with the changes, to confirm that we are not regressing either. (Just curiosity! I agree that this PR is the right direction)

[lalitb]

While there are no current perf benefits with the approach,

Can you post benchmarks/stress tests with the changes, to confirm that we are not regressing either. (Just curiosity! I agree that this PR is the right direction)

Have added the stress and benchmark results in the description. Surprisingly I see the improvement, but don't want to read much out of that as of now :)

[cijothomas]

wondering why is this optional?

[lalitb]

Oh, this was existing code in API. I can have a look into this separately.

[cijothomas]

LGTM.
Waiting to add changelog, and remove the API to set span/trace context before approving.

[lalitb]

LGTM. Waiting to add changelog, and remove the API to set span/trace context before approving.

changelog is added now. Thanks for the review.

[cijothomas]

Thanks! Great to see API getting leaner and delegating things to the SDK!

[cijothomas]

@TommyCpp could you take another look?

@lalitb the CI is red, I think it is related to doc failure in this pr. can you fix that?

[cijothomas]

@lalitb also - can you update PR desc. to reflect the final change in the PR? (Set vs Add and anything else) Thank you!

[lalitb]

@lalitb the CI is red, I think it is related to doc failure in this pr. can you fix that?

done

@lalitb also - can you update PR desc. to reflect the final change in the PR? (Set vs Add and anything else) Thank you!

done.

[TommyCpp]

Looks good. Thanks for the change