[
https://issues.apache.org/jira/browse/HBASE-2248?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12841669#action_12841669
]
Todd Lipcon commented on HBASE-2248:
------------------------------------
Hey Ryan
I looked over this patch a bit this afternoon. It's clever but I think it can
result in loss of read-your-own-writes consistency for a single client.
Consider this scenario:
|| Action || Read # || Write # || memstoreRead || memstoreWrite ||
| Client A begins a put on row R | - | 1 | 0 | 1 |
| Client B begins a put on row S | - | 2| 0 | 2 |
| Client B finishes a put on row S | - | - | 0 | 2 |
| Client B initiates a get on row S | 0 | - | 0 | 2 |
So, since client A's put #1 is still ongoing on a separate row, client B is
unable to read version #2 of its row.
I think dropping consistency below read-your-own-writes is bad, even though
it's rare that the above situation would occur. Under high throughput I think
it's possible to occur, and it could be very very bad if people are relying on
this level of consistency to implement transactions, etc.
One possible solution is that completeMemstoreInsert can spin until
memstoreRead >= e.getWriteNumber(). Given that it only has to wait for other
concurrent writers to finish, a spin on memstoreRead.get() should only go a few
cycles and actually be reasonably efficient.
I'll think a bit about whether there are other possible solutions.
> Provide new non-copy mechanism to assure atomic reads in get and scan
> ---------------------------------------------------------------------
>
> Key: HBASE-2248
> URL: https://issues.apache.org/jira/browse/HBASE-2248
> Project: Hadoop HBase
> Issue Type: Bug
> Affects Versions: 0.20.3
> Reporter: Dave Latham
> Fix For: 0.20.4
>
> Attachments: HBASE-2248-demonstrate-previous-impl-bugs.patch,
> HBASE-2248-ryan.patch, hbase-2248.gc, HBASE-2248.patch, Screen shot
> 2010-02-23 at 10.33.38 AM.png, threads.txt
>
>
> HBASE-2037 introduced a new MemStoreScanner which triggers a
> ConcurrentSkipListMap.buildFromSorted clone of the memstore and snapshot when
> starting a scan.
> After upgrading to 0.20.3, we noticed a big slowdown in our use of short
> scans. Some of our data repesent a time series. The data is stored in time
> series order, MR jobs often insert/update new data at the end of the series,
> and queries usually have to pick up some or all of the series. These are
> often scans of 0-100 rows at a time. To load one page, we'll observe about
> 20 such scans being triggered concurrently, and they take 2 seconds to
> complete. Doing a thread dump of a region server shows many threads in
> ConcurrentSkipListMap.biuldFromSorted which traverses the entire map of key
> values to copy it.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
