Every operation by transferring value between memcache and datastore is a non atomic operation. The original design is cache value from datastore to memcache. The equivalent implementation is just throwing random exception from the transaction.
I feel the non sharded counter plus a memcache will be more robust solution. Because a reasonable frequency update from memcache to datastore completely eliminated the necessity of sharded implementation. A high frequency memcache to datastore implementation will be very unreliable due to high frequency non atomic operations. On Nov 11, 4:12 pm, Bill <[EMAIL PROTECTED]> wrote: > I probably should've piped up on the thread earlier. I'm currently > looking at yejun's fork and will merge pending some questions I have > on his optimizations. > > Here's one: > My old code stored the counter name in each shard so I could get all > shards with a single fetch. If you have 20 shards, you could have any > number of actually created shards. In a very high transaction system, > probably all 20 shards exist. > In yejun's optimization, he's iterating through each shard using > get_by_key_name and checking if it exists. Which is likely to be > faster? > > A nice optimization by yejun is making count a TextProperty. This > will prevent indexing and would probably save some cycles. > > Josh, you said "lines 144-145 should be under the try: on line 137. > That way, the delayed counter count won't get reset to zero even in > the case of a failed transaction." > > I thought any datastore errors are handled via the db.Error exception > which forces a return before the delayed counter count is reset. > (db.Error is defined in appengine/api/datastore_errors.py) > > On Josh's memcached buffering scheme, I can definitely see the utility > if you're willing to sacrifice some reliability (we're down to one > point of failure -- memcache) for possibly a lot of speed. Using > memcache buffers for counters makes sense because it's easy to > accumulate requests while for other model puts, like comments or other > text input, the amount of memcache buffer could grow pretty large > quickly. > > Does it make sense to use a sharded backend to a memcache buffer? > Depends on the frequency of the final datastore writes, as mentioned > above. (I'm not as concerned with complexity as yejun because I think > this buffering is reasonably simple for counters.) So I think this > would be a good thing to add onto the sharded counter code through a > fork. Then people who want that kind of speed can opt for it. > > -Bill --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "Google App Engine" group. To post to this group, send email to google-appengine@googlegroups.com To unsubscribe from this group, send email to [EMAIL PROTECTED] For more options, visit this group at http://groups.google.com/group/google-appengine?hl=en -~----------~----~----~----~------~----~------~--~---