"Set" and then "get" don't retrieve the stored value

Brad Fitzpatrick brad at danga.com
Mon Nov 20 17:12:14 UTC 2006


On Mon, 20 Nov 2006, Don MacAskill wrote:

>
> Oh, interesting.  That might explain some strange behavior we were
> seeing.  I guess I'll try adding a sleep(1) after a flush_all, prior to
> enabling the server again, and see if that fixes our problem.
>
> Wish I had a clever way of solving the base problem though.  Alas, I
> don't.  :)

Yeah...

We could do gross stuff like using the lower (always zero) bits of various
pointers in item* for the world generation number, but I think the CPU
cost of maintain those wouldn't be worth it.




>
> Don
>
>
> Brad Fitzpatrick wrote:
> > This is another instance of the flush_all-is-only-second-granularity
> > bug that's existed for awhile.
> >
> > Because you flush_all and then do a set within the rest of that second,
> > those sets are also considered non-existent.
> >
> > We've wanted to fix this by adding a world generation number (and storing
> > the world version of each item in the item struct), or making the time
> > higher resolution, but both increase the size of *item for little gain.
> >
> > Anybody have clever ideas on how to fix this?
> >
> > (btw, there's already a failing TODO test for this in the test suite, in
> > the flush-all.t test....)
> >
> >
> >
> > On Tue, 14 Nov 2006, Kristian Hellquist wrote:
> >
> >> Here's the dump. The key to look for is
> >> '65b05a7489ea704c5afb80fee7b029497a80beac'. As you can see I do a
> >> "set" with response "STORED" followed by a delete which returns "NOT
> >> FOUND". After that follows various calls that I think you can ignore.
> >>
> >> Thanks for your help,
> >> Kristian
> >>
> >>
> >> 2006/11/14, Brad Fitzpatrick <brad at danga.com>:
> >>> I'm coming into this late.
> >>>
> >>> Why would a "get" right after a "set" not work?  This isn't a known issue,
> >>> at least not to me.
> >>>
> >>> If you can reproduce this easily, send me a tcpdump pcap file of the
> >>> traffic to your memcached.
> >>>
> >>> If you're using the loopback to a 127.0.0.1 memcached on port 11211, run
> >>> this:
> >>>
> >>> # tcpdump -w forbrad.pcap -s 0 -i lo port 11211
> >>>
> >>> Then do your (minimal!) failing test, control-C the tcpdump, and email me
> >>> the forbrad.pcap file.
> >>>
> >>>
> >>> On Tue, 14 Nov 2006, Kristian Hellquist wrote:
> >>>
> >>>> The problem is that the the small window unfortunately seems to be
> >>>> quite large for me. Do you have measure in time of how small the
> >>>> window is?
> >>>>
> >>>> Basically, I have some unit tests on a web application. The web
> >>>> application relies on memcached heavily. Everything with memcached has
> >>>> worked like a charm for us so far. We cache sql-finders in memcached
> >>>> and invalidates the sql-finders when the model has changed.
> >>>>
> >>>> In a specific use-case we dont want to update the state of the db
> >>>> immediately. This data change *a lot*. So when the state of model
> >>>> change we will cache the sql-finder in memcache directly and later on
> >>>> we flush the change to the db.  It is in this use-case that I have
> >>>> experienced the "set" and "get" problem.
> >>>>
> >>>> My unit-test successfully ran when I added a sleep call of 0.3 s
> >>>> between each memcache-call :/ This is a lot. Is there a way of
> >>>> configurate memcached so I experience greater performance?
> >>>>
> >>>> I know that the test-environment doesnt mimick the user behavior of
> >>>> the webapplication in real use. We use rails and the unit tests
> >>>> doensn't involve the webserver at all. So in real life more delay will
> >>>> be inserted between each "set" and "get" call. But, can I *trust*
> >>>> memcache that data has been stored if I receive a "STORED" response?
> >>>> Is just a matter of time before the value can be retrieved again? Or
> >>>> is it bad usage of memcache to cache the sql-finders directly and
> >>>> later on update the db?
> >>>>
> >>>> Thanks for your reply,
> >>>> Kristian Hellquist
> >>>>
> >>>>
> >>>> 2006/11/14, Jehiah Czebotar <jehiah at gmail.com>:
> >>>>>> Basically, yes.  There is a small window where this can happen.  There
> >>>>>> have been numerous discussions about it on the list.  Most people note
> >>>>>> it with a set and then delete.
> >>>>>>
> >>>>> tif my memory serves me, the correct order is a delete, set, then a
> >>>>> get all for the same key
> >>>>>
> >>>>
> >
>
>


More information about the memcached mailing list