fix: #2033 shared queue push bug

[keroxp]

• fixed Uncaught Error: Expecting promise in table. 0 #2033

• I finally found SharedQueue (js, also rust) wrongly push record item though it has full records according to MAX_RECORDS size.

• shifting operation to SharedQueue that pushed more than MAX_RECORDS will cause obtaining wrong subarray of shared bytes.

• That is why deserialized flatbuffers message has cmd_id=0

• NOTE: this fixing seriously degrade benchmarks

• benchmark comparison between https://deno.land/std/http/http_bench.ts is below

[email protected] (ended with error #2033 )

 ➜  wrk -t12 -c400 -d30s http:https://127.0.0.1:4500
Running 30s test @ http:https://127.0.0.1:4500
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    18.86ms    3.58ms  62.30ms   86.99%
    Req/Sec     1.67k   203.47     2.77k    89.08%
  597492 requests in 30.02s, 28.49MB read
  Socket errors: connect 0, read 383, write 0, timeout 0
Requests/sec:  19904.28
Transfer/sec:      0.95MB

this PR (no error after benching)

➜  wrk -t12 -c400 -d30s http:https://127.0.0.1:4500
Running 30s test @ http:https://127.0.0.1:4500
  12 threads and 400 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    99.21ms   12.93ms 262.61ms   84.30%
    Req/Sec   315.41     45.03   564.00     87.41%
  112963 requests in 30.10s, 5.39MB read
  Socket errors: connect 0, read 451, write 0, timeout 0
Requests/sec:   3753.16
Transfer/sec:    183.26KB

I don't fully understand how SharedQueue is needed, MAX_RECORDS seems to be small for intensive use cases...?

[ry]

Cool nice work. I’ve only given this a quick skim - but it’s not clear to me why there’s a performance degradation - do you have an idea? this test demonstrates the problem you were seeing (it fails before the fix is applied)?

[keroxp]

No idea... I'm investigating around here. https://github.com/denoland/deno/blob/master/core/isolate.rs#L453

[ry]

@keroxp I suspect the performance hit might be due to the debug logging. I've minimized your changes into a smaller patch for perf testing here: #2158
(Please see if this is still fixing your bug)

[ry]

closed in favor of #2158