https://pinot.apache.org/ logo
Join Slack
Powered by
# pinot-helix
  • k
    may be its bcos of mac
  • k
    let me restart
  • j
    I also run it on mac
  • j
    Copy code
    3.3 GHz Intel Core i7
16 GB 2133 MHz LPDDR3
  • j
    Testing it on Linux now
  • j
    Total time to spend in local Linux box: Latest Version:
    Copy code
    Total time: 13735
Total time: 13970
Total time: 14209
    Pinot 1.0.0:
    Copy code
    Total time: 13273
Total time: 12773
Total time: 12436
  • j
    The query execution part is always under 600ms (i.e. after printing 
    ***** Offline quickstart setup complete *****
    )
  • k
    Okay. I will debug why it’s taking more time on my Mac. Let’s continue with this.
  • k
    So new code is 3 seconds slower than previous version
  • k
    That’s 20%
  • k
    Where is that slowness coming from
  • j
    I added the log messages in each of the steps and here are the stats: Pinot 1.0.0:
    Copy code
    Start components: 5504
schema update time: 398
add table time: 103
Finish building segment time: 2118
Finish segment push time: 377
Total setup time: 13920
    Copy code
    Start components: 5978
schema update time: 364
add table time: 98
Finish building segment time: 2116
Finish segment push time: 392
Total setup time: 14266
    Copy code
    Start components: 5407
schema update time: 463
add table time: 183
Finish building segment time: 2071
Finish segment push time: 406
Total setup time: 13813
    Copy code
    Start components: 5001
schema update time: 369
add table time: 109
Finish building segment time: 2314
Finish segment push time: 451
Total setup time: 13616
  • j
    Latest version:
    Copy code
    Start components: 7054
schema update time: 455
add table time: 130
Finish building segment time: 2673
Finish segment push time: 706
Total setup time: 16151
    Copy code
    Start components: 6553
schema update time: 392
add table time: 122
Finish building segment time: 2499
Finish segment push time: 435
Total setup time: 15120
    Copy code
    Start components: 6618
schema update time: 468
add table time: 148
Finish building segment time: 2433
Finish segment push time: 624
Total setup time: 15425
    Copy code
    Start components: 6938
schema update time: 565
add table time: 152
Finish building segment time: 2906
Finish segment push time: 464
Total setup time: 16156
  • j
    Tested the time to spend when starting all the components. It seems the latest Helix version took more time on starting the ZK.
  • j
    Latest version:
    Copy code
    Start ZK: 1690
Start controller: 4699
Start broker: 491
Start server: 546
Total time to start all components: 7426
    Copy code
    Start ZK: 1733
Start controller: 6018
Start broker: 507
Start server: 640
Total time to start all components: 8898
    Copy code
    Start ZK: 1072
Start controller: 3930
Start broker: 539
Start server: 633
Total time to start all components: 6174
    Copy code
    Start ZK: 1050
Start controller: 3772
Start broker: 417
Start server: 503
Total time to start all components: 5742
    Copy code
    Start ZK: 1271
Start controller: 4366
Start broker: 451
Start server: 560
Total time to start all components: 6648
  • j
    Pinot 1.0.0:
    Copy code
    Start ZK: 306
Start controller: 3954
Start broker: 499
Start server: 472
Total time to start all components: 5231
    Copy code
    Start ZK: 487
Start controller: 5781
Start broker: 580
Start server: 447
Total time to start all components: 7295
    Copy code
    Start ZK: 461
Start controller: 6659
Start broker: 606
Start server: 426
Total time to start all components: 8152
    Copy code
    Start ZK: 283
Start controller: 3946
Start broker: 466
Start server: 501
Total time to start all components: 5196
    Copy code
    Start ZK: 280
Start controller: 3708
Start broker: 435
Start server: 414
Total time to start all components: 4837
  • j
    The start time of broker and server seem to be the same. For controller, the main difference is that latest version registers Pinot controller as Helix controller + Helix participant, whereas Pinot 1.0.0 registers as Helix controller only.
  • j
    @User So after bumping up Helix version, it takes longer time on starting ZK. And there's no code changes on starting ZK in Pinot codebase between these two versions.
  • j
    You mean start ZK in the tests?
  • j
    Yes
  • j
    The only thing related might be we upgraded ZK to 3.4.13 for node DNS update.
  • j
    Before, it is 3.4.11
  • j
    Just to confirm, starting ZK is not related to Helix logic, right? And we don't need to do it in a real production environment.
  • j
    @User do you remember the reason why we required Helix team to upgrade ZK client?
  • j
    @User That is because we want the participant dynamically update the ZK server ips when DNS is changed.
  • s
    We had to upgrade zk client because the old zk client had a "feature" that resolved the dns once and cached the IP addresses of zookeeper. This meant that if we upgrade zookeeper servers, the entire cluster (or, at least the participants connected to the decommissioned zk) had to be restarted. In order to be able to roll out and roll in new zk servers, we wanted that the zk client automatically re-resolve IP addresses if they fail to connect. The helix team came up with the recommendation that upgrading zk client will fix that problem. Hence we upgraded the zk client.
  • k
    It’s fine if startup is slow because of ZK startup since this won’t be a problem in production
  • k
    Thanks for the investigation Jack. I will review the PR to avoid NPE in QuickStart
    👍 1
  • j
    @User ^
  • h
    hi @User