[
https://issues.apache.org/jira/browse/HBASE-5416?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13215720#comment-13215720
]
Nicolas Spiegelberg commented on HBASE-5416:
--------------------------------------------
Overall, I agree that this is a useful design pattern. We use this pattern in
our messages deployment and other production use cases as well. I'm more
concerned about this being in the critical path. This is deep in the core
logic, which has a lot of complicated usage and is extremely bug-prone (even
after extensive unit tests).
If you don't need atomicity, then you don't get much benefit from solving this
in the critical path. The change introduces a lot of risk and design decisions
that we have to worry about years later. It might be some work to understand
how to use a batch factor; but don't you think it would take more work to
understand the variety of use cases for scans to ensure that we don't introduce
side effects and make a scalable architectural decision?
At the very least, we should get a scan expert to look at this code before
committing. I'm not one, but I know this isn't the same as making a business
logic change. I just have one question about the patch right now: Should we
have unit tests case for ensuring the interop between this feature and 'limit'?
For example, ensure that joinedHeap is scanned before going to the next row if
the storeHeap results.size() == limit
> Improve performance of scans with some kind of filters.
> -------------------------------------------------------
>
> Key: HBASE-5416
> URL: https://issues.apache.org/jira/browse/HBASE-5416
> Project: HBase
> Issue Type: Improvement
> Components: filters, performance, regionserver
> Affects Versions: 0.90.4
> Reporter: Max Lapan
> Assignee: Max Lapan
> Attachments: 5416-v5.txt, 5416-v6.txt, Filtered_scans.patch,
> Filtered_scans_v2.patch, Filtered_scans_v3.patch, Filtered_scans_v4.patch
>
>
> When the scan is performed, whole row is loaded into result list, after that
> filter (if exists) is applied to detect that row is needed.
> But when scan is performed on several CFs and filter checks only data from
> the subset of these CFs, data from CFs, not checked by a filter is not needed
> on a filter stage. Only when we decided to include current row. And in such
> case we can significantly reduce amount of IO performed by a scan, by loading
> only values, actually checked by a filter.
> For example, we have two CFs: flags and snap. Flags is quite small (bunch of
> megabytes) and is used to filter large entries from snap. Snap is very large
> (10s of GB) and it is quite costly to scan it. If we needed only rows with
> some flag specified, we use SingleColumnValueFilter to limit result to only
> small subset of region. But current implementation is loading both CFs to
> perform scan, when only small subset is needed.
> Attached patch adds one routine to Filter interface to allow filter to
> specify which CF is needed to it's operation. In HRegion, we separate all
> scanners into two groups: needed for filter and the rest (joined). When new
> row is considered, only needed data is loaded, filter applied, and only if
> filter accepts the row, rest of data is loaded. At our data, this speeds up
> such kind of scans 30-50 times. Also, this gives us the way to better
> normalize the data into separate columns by optimizing the scans performed.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
