Use mapChunks in CirceInstances.streamedJsonArray

[peterneyens]

Small clean up to use mapChunks instead of an implementation using repeatPull.

[danicheg]

Really like this refactoring 👍🏻 but wouldn't you mind adding a small micro-benchmark to the CirceJsonBench suite to prove we aren't losing any microseconds with that implementation?

[diesalbla]

Perhaps this can be a local def nonFirst(c: Chunk[Json]): Chunk[Byte] function?

[diesalbla]

Additionally, rather than using the fromJsonToChunk on every element, and using chunk concatenation (in the bldr builder), would it also help to use one bytebuffer per source chunk, directly print each json to that buffer, and use one view ?

Edit perhaps that would be an optimisation to do at the circe library instead, with a printInterspersed or printCommaSeparated method that takes any iterable of Json, or similar. https://github.com/circe/circe/blob/6bbb7e73a529916faec89f1455d11356af03ea64/modules/core/shared/src/main/scala/io/circe/Printer.scala#L209-L221.

[peterneyens]

Additionally, rather than using the fromJsonToChunk on every element, and using chunk concatenation (in the bldr builder), would it also help to use one bytebuffer per source chunk, directly print each json to that buffer, and use one view ?

Benchmarking streamJsonArrayEncoder and jsonEncoder makes it clear that there is a lot of room for improvement. That feels a bit out of scope of my small refactoring in this PR.

[peterneyens]

Finally found some time to add a benchmark.

Previous

[info] Benchmark                           (elems)  (elemsPerChunk)  Mode  Cnt        Score        Error  Units
[info] CirceJsonStreamBench.encode_stream      100               50  avgt    5    54294.630 ±  13951.153  ns/op
[info] CirceJsonStreamBench.encode_stream      100              500  avgt    5    50007.465 ±   2632.255  ns/op
[info] CirceJsonStreamBench.encode_stream     1000               50  avgt    5   299525.990 ±  54286.551  ns/op
[info] CirceJsonStreamBench.encode_stream     1000              500  avgt    5   251252.929 ± 101122.138  ns/op
[info] CirceJsonStreamBench.encode_stream    10000               50  avgt    5  2731966.198 ± 760253.727  ns/op
[info] CirceJsonStreamBench.encode_stream    10000              500  avgt    5  2224727.347 ± 327952.214  ns/op

This PR:

[info] Benchmark                           (elems)  (elemsPerChunk)  Mode  Cnt        Score        Error  Units
[info] CirceJsonStreamBench.encode_stream      100               50  avgt    5    52789.592 ±   7598.824  ns/op
[info] CirceJsonStreamBench.encode_stream      100              500  avgt    5    49110.386 ±   1697.552  ns/op
[info] CirceJsonStreamBench.encode_stream     1000               50  avgt    5   280540.460 ±  30941.520  ns/op
[info] CirceJsonStreamBench.encode_stream     1000              500  avgt    5   237874.710 ±   6312.693  ns/op
[info] CirceJsonStreamBench.encode_stream    10000               50  avgt    5  2564971.598 ±  92078.114  ns/op
[info] CirceJsonStreamBench.encode_stream    10000              500  avgt    5  2214598.872 ± 685056.424  ns/op

[diesalbla]

Some benchmarks often use a "Throughput" mode, rather than average. Perhaps that would help smooth out results error?