Delete Objects based on REGEX.

[Antonello Provenzano]

Steven,

> I actually implemented regex deletes quite some time back. But the
> problem is that I did it in a brute-force way, nothing elegant or
> efficient, and if you have a large cache the whole thing will freeze up
> for many seconds, chewing 100% of your CPU while the code walks through
> all the items comparing them against the regex. We wanted it for very
> occasional administrative operations where the cache freezing up for a
> short while would be less disruptive than a full flush, but I think
> we've used it maybe two or three times at most because it's so nasty.

That's what I was trying to say. I created a project for Mono project
(named Deveel Regex) for the compilation and execution of RegEx
expressions on strings and I know what I'm speaking about: although
Deveel Regex is quite fast (faster than the built in solution in
Microsoft .NET and Mono), the parsing, compilation and execution of a
RegEx is a quite stressing process, specially if repeated thousands of
times per second.

Maybe the amount of memory and CPU used by Perl is much smaller than
Mono or .NET (which are behind a virtual machine as Java), since I've
never worked with it I have no benchmarks for this, I still believe
that RegEx in this kind of context could be a killing feature that
would stress the machine memcached would run on and cause problems in
long terms.

Antonello



On 7/26/07, Steven Grimm <sgrimm at facebook.com> wrote:
> I actually implemented regex deletes quite some time back. But the
> problem is that I did it in a brute-force way, nothing elegant or
> efficient, and if you have a large cache the whole thing will freeze up
> for many seconds, chewing 100% of your CPU while the code walks through
> all the items comparing them against the regex. We wanted it for very
> occasional administrative operations where the cache freezing up for a
> short while would be less disruptive than a full flush, but I think
> we've used it maybe two or three times at most because it's so nasty.
>
> Brad's generation-based idea makes much more sense as a long-term
> approach to this feature, but honestly I like tags a lot better. I'd
> rate tags a 10 and regex deletes, even well-implemented ones, at most a
> 6 or 7 because they're so inflexible and require that all consumers of
> objects in the cache be aware of all the metadata you can use to address
> a particular object (because it's all part of the key and you need the
> whole key to query the cache.) Tags are IMO far superior because you can
> tag an item with metadata that some of your clients are blissfully
> unaware of.
>
> -Steve
>
>
> BUSTARRET, Jean-francois wrote:
> > IMHO, regexp deletes would be great. I would prefer tags, but on a scale of 10 (0 being the current state), tags would be at 10, and regexps at 9.
> >
> > I my case, it would solve the most important cache coherency problems.
> >
> > I'm willing to help (I would be unable to code the most complex part of it, but still willing to help).
> >
> >
> >> -----Message d'origine-----
> >> De : memcached-bounces at lists.danga.com
> >> [mailto:memcached-bounces at lists.danga.com] De la part de
> >> Dustin Sallings
> >> Envoyé : mercredi 25 juillet 2007 23:14
> >> À : dormando
> >> Cc : P R E M; memcached at lists.danga.com
> >> Objet : Re: Delete Objects based on REGEX.
> >>
> >>
> >> On Jul 25, 2007, at 12:59 , dormando wrote:
> >>
> >>
> >>> Seems interesting, don't recall if anyone was specifically
> >>>
> >> against it.
> >>
> >>> The implementation details are a bit foggy, and someone has
> >>>
> >> to go do
> >>
> >>> it still.
> >>>
> >>      The biggest complaints are that people feel that it may
> >> not solve actual problems people have.  The alternative is
> >> tags, but that seems more complicated (to use) to me.
> >>
> >>      In the end, if nobody's problems are solved, people
> >> probably aren't going to be writing any code for it.
> >>
> >> --
> >> Dustin Sallings
> >>
> >>
> >>
> >>
>
>
>