Structured logging and structured concurrency are a match made in heaven. Using task locals for loggers enables easier passing of metadata for log messages and removes the cumbersome passing around of loggers.
Modification
This PR adds a proposal, the implementation and benchmarks for a task local logger.
Result
Using a task local for logging makes it incredible ergonomic to carry around metadata. To be transparent the benchmarks show that this is around 20x slower than a bare no-op log. A no-op log takes around 2ns whereas a task local logger no-op log takes around 40ns. 95% of that time is spent in accessing the task local since that is missing some @inlinable in the runtime and goes through the C interface. Instruction wise it is 20 vs 290 instructions.
Now I still propose that we go forward with this since the ergonomics far outweigh the performance initially. Furthermore, this acts as a great forcing function for performance improvements in the runtime since logging is often done in hot paths.
Motivation
Structured logging and structured concurrency are a match made in heaven. Using task locals for loggers enables easier passing of metadata for log messages and removes the cumbersome passing around of loggers.
Modification
This PR adds a proposal, the implementation and benchmarks for a task local logger.
Result
Using a task local for logging makes it incredible ergonomic to carry around metadata. To be transparent the benchmarks show that this is around 20x slower than a bare no-op log. A no-op log takes around 2ns whereas a task local logger no-op log takes around 40ns. 95% of that time is spent in accessing the task local since that is missing some
@inlinablein the runtime and goes through the C interface. Instruction wise it is 20 vs 290 instructions.Now I still propose that we go forward with this since the ergonomics far outweigh the performance initially. Furthermore, this acts as a great forcing function for performance improvements in the runtime since logging is often done in hot paths.