[jira] [Updated] (KUDU-2939) built-in NTP client: implement stricter source selection algorithm

["Alexey Serbin (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/KUDU-2939?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Alexey Serbin updated KUDU-2939:
--------------------------------
        Parent: KUDU-2935
    Issue Type: Sub-task  (was: Bug)

> built-in NTP client: implement stricter source selection algorithm
> ------------------------------------------------------------------
>
>                 Key: KUDU-2939
>                 URL: https://issues.apache.org/jira/browse/KUDU-2939
>             Project: Kudu
>          Issue Type: Sub-task
>            Reporter: Alexey Serbin
>            Priority: Major
>
> In the initial implementation, the built-in NTP client has a simplified NTP source selection/filterting algorithm: basically, it's about sanitisation of incoming NTP packets and ignoring falsetickers.  It's necessary to make the source selection algorithm more robust.
> At least, it's necessary to filter out NTP servers with higher than allowed root distance and jitter.
> Also, as a part evaluation of the quality of a source (i.e. NTP server) in the source selection algorithm, it's necessary to introduce corresponding knobs (via gflags) to set the thresholds for the following properties of a source:
> * Maximum distance of a source (i.e. NTP server) from the root.  The distance also includes the accumulated dispersion.  Servers which are no longer synchronized might get their root distance grow high, and having this threshold is a way to avoid synchronising with servers which have contacted their own servers too long ago.  By default, the maximum root distance should be set to 3 seconds (in the initial implementation, it's 16 seconds -- the maximum allowed by the RFC itself).
> * Maximum jitter (i.e. standard deviation) of time samples. This should prevent synchronisation with sources that have a small root distance, but their time is too variable.  By default, this should be set to 1 second.



--
This message was sent by Atlassian Jira
(v8.3.2#803003)