Hi!
I've applied this patch. It should be pushed within the hour.
Thanks,
-Brian
On Apr 14, 2009, at 3:50 AM, Nikhils wrote:
Hi Brian,
A careful perusal of the memcached_mget_by_key() function revealed
the bug.
The code was generating the serverid by using the master key first.
However if the serverid was 0, the following check was causing re-
evaluation of the serverid by using the buffer key leading to the
query being sent to another server:
if (master_server_key) /* if the serverid had evaluated to zero,
we go to else */
server_key= master_server_key;
else
server_key= memcached_generate_hash(ptr, keys[x],
key_length[x]);
After fixing this, the *_by_key logic appears to be working just fine.
Please find attached, a patch (diff -c format) which fixes this.
Regards,
Nikhils
On Tue, Apr 14, 2009 at 12:27 PM, Nikhils <[email protected]> wrote:
Hi,
I recently shifted to using the memcached_get/set_by_key APIs. The
trouble is that even after warming the cache with all of the
content, a subsequent query to fetch the key/value pair fails with a
not found error. I am seeing this behaviour in a 2-node memcached
cluster.
Its occurring consistently and on reverting back to using the normal
memcached_get/set APIs, the key/value pairs can be retrieved
appropriately! However I do need to use these by_key APIs for my
data consistency logic.
I have added appropriate logging information in my C code (am
linking to libmemcached 0.27) and confirmed that the master key is
the same in both set_by_key/get_by_key calls.
Anything obvious that I am missing? Has anyone observed such dubious
behaviour before with these by_key calls? Comments appreciated.
Regards,
Nikhils
<libmemcached_mget_by_key_bug.patch>
--
_______________________________________________________
Brian "Krow" Aker, brian at tangent.org
Seattle, Washington
http://krow.net/ <-- Me
http://tangent.org/ <-- Software
_______________________________________________________
You can't grep a dead tree.
