Michael Han created ZOOKEEPER-2627:
--------------------------------------
Summary: Remove ZRWSERVERFOUND from C client and replace
handle_error with something more semantically explicit for r/w server reconnect.
Key: ZOOKEEPER-2627
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2627
Project: ZooKeeper
Issue Type: Bug
Components: c client
Affects Versions: 3.5.2
Reporter: Michael Han
Assignee: Michael Han
Fix For: 3.5.3
While working on ZOOKEEPER-2014, I noticed a discrepancy between Java and C
client regarding the error codes definition. There is a
{noformat}ZRWSERVERFOUND = -122{noformat} definition in C client which is not
present in Java client's KeeperException.Code definitions.
This discrepancy was introduced by ZOOKEEPER-827, where the C client logic was
simulating the Java client's logic when doing a read/write server search while
client is in read only mode. Once client finds a valid read/write server,
client will try to disconnect and reconnect with this read/write server, as we
always prefer r/w server in ro mode. The way Java client is doing this
disconnect/reconnect process is by throwing a RWServerFoundException (instead
of a KeeperException) to set the client in disconnected state, then wait for
client reconnect with r/w server address set before throwing the exception. C
client did similar but instead of having an explicitly disconnect / clean up
routine, the client was relying on handle_error to do the job where
ZRWSERVERFOUND was introduced.
I propose we remove ZRWSERVERFOUND error code from C client and use an explicit
routine instead of handle_error when we do r/w server search in C client for
two reasons:
* ZRWSERVERFOUND is not something ZK client users would need to know. It's a
pure implementation detail that's used to alter the connection state of the
client, and ZK client users have no desire nor need to handle such errors, as
R/W server scanning and connect is handled transparently by ZK client library.
* To maintain consistency between Java and C client regarding error codes
definition. Without removing this from C client, we would need to replace
RWServerFoundException in Java client with a new KeeperException, and again
with the reason mentioned above, we don't need a KeeperException for this
because such implementation detail does not have to be exposed to end users
(unless, we provided alternative for users to opt-out automate R/W server
switching when in read only mode which we don't.).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)