[
https://issues.apache.org/jira/browse/BOOKKEEPER-1102?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sijie Guo resolved BOOKKEEPER-1102.
-----------------------------------
Resolution: Fixed
Fix Version/s: 4.5.0
Issue resolved by merging pull request 275
[https://github.com/apache/bookkeeper/pull/275]
{noformat}
commit a5f8580f53464065243a9af038935f5893434166
Author: Samuel Just <[email protected]>
AuthorDate: Mon Jul 31 23:34:07 2017 -0700
Commit: Sijie Guo <[email protected]>
CommitDate: Mon Jul 31 23:34:07 2017 -0700
BOOKKEEPER-1102: Clarify BookieInfoReader and fix associated test flappers
BookieInfoReader:
The previous syncronization logic wasn't really correct, and the logic
at the top of the method was far more complicated than it needed to be.
Restrict bookies to be non-null. Restructure the code to simply use
the BookieInfoReader instance as a single lock.
One significant behavioral change is that we scan every bookie not in
the map, and we clear from the map bookies which returned an error.
Also, explicitely cache the most recent bookie set reported by the
BookieWatcher. This eliminates the need to call into BookieWatcher
from getReadWriteBookieInfo and the corresponding error path. The
periodic scan continues to explicitely check.
Another departure is the addition of an explicit retry-on-error param to
trigger retry if any of the requests failed
(getBookieInfoRetryIntervalSeconds). We'll only retry the ones that
actually failed (along with any new additions since the last run). This
is useful because bookie startup triggers the addition of the bookie
node to zk before the bookie actually becomes available for the bookie
info request, so there can be rare races in the unit tests between
BookieInfoReader requesting the info and the bookie actually being up.
Also, add a method to allow tests to wait for updates to be reflected.
PerChannelBookieClient: fix error handling for BookieInfo keys
Passing a key corresponding to a GET_BOOKIE_INFO operation to
errorOutReadKey results in a casting exception, clean up the invalid
calls.
BookKeeperClusterTestCase: add killBookieAndWaitForZK
Should reduce the need for tests to wait for an arbitrary period to let
the cluster "settle".
BookKeeperDiskSpaceWeightedLedgerPlacementTest:
This test was heavily time dependent, and the Thread.sleep values did
not work universally. Instead, eliminate the arbitrary Thread.sleep
values and instead verify the free space changes on each change.
Also, switch the delay on
testDiskSpaceWeightedBookieSelectionWithPeriodicBookieInfoUpdate
to simply use an atomic boolean to signal the value switch.
Signed-off-by: Samuel Just <sjustsalesforce.com>
Author: Samuel Just <[email protected]>
Reviewers: Enrico Olivelli <[email protected]>
This patch had conflicts when merged, resolved by
Committer: Sijie Guo <[email protected]>
This closes #275 from athanatos/forupstream/BOOKKEEPER-1102
{noformat}
> org.apache.bookkeeper.client.BookKeeperDiskSpaceWeightedLedgerPlacementTest.testDiskSpaceWeightedBookieSelectionWithBookiesBeingAdded
> is unreliable
> ---------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: BOOKKEEPER-1102
> URL: https://issues.apache.org/jira/browse/BOOKKEEPER-1102
> Project: Bookkeeper
> Issue Type: Bug
> Components: bookkeeper-client
> Affects Versions: 4.5.0
> Reporter: Samuel Just
> Assignee: Samuel Just
> Priority: Minor
> Fix For: 4.5.0
>
>
> org.apache.bookkeeper.client.BookKeeperDiskSpaceWeightedLedgerPlacementTest.testDiskSpaceWeightedBookieSelectionWithBookiesBeingAdded
> can intermittently fail depending on the timing of the client receiving the
> info back from the bookies.
> Additionally, the synchronization in BookieInfoReader is more complicated
> than necessary and not entirely correct.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
