Prometheus Write Request Deserialization Benchmark

This repository serves as an example for pursuing better deserialization performance of the well-known Protobuf-encoded Prometheus WriteRequest.

As a baseline, the deserialization implementation generated by prost takes about 7.3ms to decode a WriteRequest with 10k timeseries on our test environment.

This repository contains several branches that each did some optimization effort.

• step1/reproduce: reproduce the baseline
• step2/repeated_field: use RepeatedField to replace Vec for repeated field to enable pooling
• step3/bytes: use bytes to facilitate zero-copy deserialization
• step4/bytes-eliminate-one-copy: eliminate one copy_to_bytes invocation
• step5/bench-bytes-slice: reproduce the overhead brought by Bytes::slice
• step6/optimize-slice: hack Bytes::slice while ensuring lifetime guarantee

Finally, we cut the deserialization cost from 7.3ms to 1.6ms.