[
https://issues.apache.org/jira/browse/KUDU-3060?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Grant Henke updated KUDU-3060:
------------------------------
Labels: roadmap-candidate (was: )
> Add a tool to identify potential performance bottlenecks
> --------------------------------------------------------
>
> Key: KUDU-3060
> URL: https://issues.apache.org/jira/browse/KUDU-3060
> Project: Kudu
> Issue Type: Improvement
> Components: CLI, perf, ui
> Reporter: Andrew Wong
> Priority: Major
> Labels: roadmap-candidate
>
> When we hear users wondering why their workloads are slower than expected,
> some common questions arise. It'd be great if we had a single tool (or a
> single webpage) that aggregated and displayed useful information for a
> specific tablet or table. Things like, for a specific table:
> - How many partitions and replicas exist for the table.
> - For those replicas, how they are distributed across tablet servers.
> - For those tablet servers, what the block cache configuration is, and what
> the current block cache stats (hit ratio, evictions, etc) are.
> - For those tablet servers, which tablets have been written to recently.
> - For those tablet servers, which tablets within the target table have been
> written to recently.
> - For those tablet servers, how many active and non-expired scanners exist.
> - For those tablet servers, which tablets within the target table have been
> read from recently.
> - For those tablet servers, how many ongoing tablet copies there are both to
> and from the server.
> - For those tablet servers, how many data directories there are.
> - For the data directories on those tablet servers, how many replicas are
> spreading data in each directory, how many blocks there are in each, and how
> much space is available in each.
> The list could go on and on. It probably makes sense to break the diagnostics
> into different phases or goals, maybe along the lines of 1) identifying
> hotspots of workloads and lag across tablet servers (e.g. a ton of writes
> going to a single tserver), and 2) digging into a single tablet server to
> understand how it's provisioned and whether that provisioning is sufficient.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
