[sheepdog] [PATCH 1/3] zookeeper: retry zk_create_seq_node on retryable error

[MORITA Kazutaka]

At Thu, 30 May 2013 22:45:37 +0800,
Liu Yuan wrote:
> 
> On 05/30/2013 10:41 PM, MORITA Kazutaka wrote:
> > But actually I've experienced many ZCONNECTIONLOSS errors by
> > zoo_create() with zookeeper 3.3.5.  I'm not sure ZOPERATIONTIMEOUT
> > also happens, though.
> 
> So what I previous meant was, zk_create() ABI is changed on newer zookeeper. I am using v3.4.5 and this is the doc in zookeeper.h
> 
>  * ZOK operation completed successfully
>  * ZNONODE the parent node does not exist.
>  * ZNODEEXISTS the node already exists
>  * ZNOAUTH the client does not have permission.
>  * ZNOCHILDRENFOREPHEMERALS cannot create children of ephemeral nodes.
>  * ZBADARGUMENTS - invalid input parameters
>  * ZINVALIDSTATE - zhandle state is either ZOO_SESSION_EXPIRED_STATE or ZOO_AUTH_FAILED_STATE
>  * ZMARSHALLINGERROR - failed to marshall a request; possibly, out of memory
>  */
> ZOOAPI int zoo_create(zhandle_t *zh, const char *path, const char *value,
>         int valuelen, const struct ACL_vector *acl, int flags,
>         char *path_buffer, int path_buffer_len);
> 
> So I guess we should add ZINVALIDSTATE in the switch case handling.

How do you think we handle ZINVALIDSTATE?

I've also experienced ZINVALIDSTATE and ZSESSIONEXPIRED, but, IIUC,
those errors are caused by a timeout.  Am I wrong?

Either way, I think the error should be handled in the different patch
unless it is a retryable one.

Thanks,

Kazutaka