Loading ...
Sorry, an error occurred while loading the content.

Re: using memcached as a StateDB.. getting there

Expand Messages
  • Thanos Chatziathanassiou
    (ignoring the fact that no-one seems interested) I came accross Cache::Memcached::Tie which pretty much does most of the work, however I m having some issues -
    Message 1 of 5 , Mar 26, 2009
    • 0 Attachment
      (ignoring the fact that no-one seems interested)

      I came accross Cache::Memcached::Tie which pretty much does most of the
      work, however I'm having some issues - actually design decisions:
      - I suppose StateDir should equal ``namespace'' in memcached parlance.
      Do we want to make this transparent to the end user or add a
      configuration option ?
      - memcached has no concept of Lock(), UnLock() and some stuff I haven't
      figured out yet. Add no-ops for these to Cache::Memcached::Tie or is
      there some more elegant way to bypass them inside Apache::ASP ?
      - It seems there's no obvious way to enumerate keys in memcached
      (FIRSTKEY NEXTKEY). Perhaps keep a separate index of keys inside
      memcached and add methods that use those (?)
      - memcached needs a few extra configuration options. Most important is
      obviously ``servers'', but also ``compress_threshold'' and ``debug''


      ---------------------------------------------------------------------
      To unsubscribe, e-mail: asp-unsubscribe@...
      For additional commands, e-mail: asp-help@...
    • Gregory S. Youngblood
      I, for one, am interested, though I haven t been able to do much or get involved. Namespace - I suggest making this an option that is user tunable, but pick
      Message 2 of 5 , Mar 27, 2009
      • 0 Attachment
        I, for one, am interested, though I haven't been able to do much or get
        involved.

        "Namespace" - I suggest making this an option that is user tunable, but pick
        a good default value that hopefully won't need to be tweaked all the time.

        "Lock/Unlock" - Each individual operation of memcached is atomic.. if two
        processes attempt to write to one location at the same time, they will be
        serialized and one will not corrupt the other. The problem is you don't
        necessarily know which one will actually win. This means you can't guarantee
        state if you don't do locking to ensure the right one wins. Consider an
        incrementing counter. Assuming you do a get to read the value, increment it
        by one, and then do a set to save it; each individual operation is safe
        (get, set), but not the combination (get+set). If two hit at the same time,
        the number would increment by 1 and not by 2 (one for each). I'd worry that
        making lock and unlock a no-op would create new problems, especially as site
        volume increases and chances of simultaneous updates increase.

        "No list of keys" - Usually I know what keys I'm stuffing into memcacheb and
        I don't need to walk through the keys, or if I do I have a place outside of
        memcache that has the keys to lookup. This one could get tricky. How would
        you keep the list of keys in a second entry intact, especially if two
        processes wanted to add a key at the same time?

        Other thoughts/suggestions:

        Having not looked at your code or design this may not be applicable, but
        consider making this generic, something that memcache or another cache
        engine could be plugged into. If you're interested, and my time permits, I'd
        be interested in working on part of this with you.

        Greg

        > -----Original Message-----
        > From: Thanos Chatziathanassiou [mailto:tchatzi@...]
        > Sent: Thursday, March 26, 2009 5:33 AM
        > Cc: 'asp@...'
        > Subject: Re: using memcached as a StateDB.. getting there
        >
        > (ignoring the fact that no-one seems interested)
        >
        > I came accross Cache::Memcached::Tie which pretty much does most of the
        > work, however I'm having some issues - actually design decisions:
        > - I suppose StateDir should equal ``namespace'' in memcached parlance.
        > Do we want to make this transparent to the end user or add a
        > configuration option ?
        > - memcached has no concept of Lock(), UnLock() and some stuff I haven't
        > figured out yet. Add no-ops for these to Cache::Memcached::Tie or is
        > there some more elegant way to bypass them inside Apache::ASP ?
        > - It seems there's no obvious way to enumerate keys in memcached
        > (FIRSTKEY NEXTKEY). Perhaps keep a separate index of keys inside
        > memcached and add methods that use those (?)
        > - memcached needs a few extra configuration options. Most important is
        > obviously ``servers'', but also ``compress_threshold'' and ``debug''
        >
        >
        > ---------------------------------------------------------------------
        > To unsubscribe, e-mail: asp-unsubscribe@...
        > For additional commands, e-mail: asp-help@...
        >
        >
        > No virus found in this incoming message.
        > Checked by AVG - www.avg.com
        > Version: 8.0.238 / Virus Database: 270.11.26/2020 - Release Date:
        > 03/25/09 07:16:00


        ---------------------------------------------------------------------
        To unsubscribe, e-mail: asp-unsubscribe@...
        For additional commands, e-mail: asp-help@...
      • Thanos Chatziathanassiou
        ... It would vaguely resemble ``StateDir . Each separate StateDir signifies a unique application. Thus far, using StateDir in shmfs/tmpfs I d use
        Message 3 of 5 , Mar 27, 2009
        • 0 Attachment
          O/H Gregory S. Youngblood έγραψε:
          > I, for one, am interested, though I haven't been able to do much or get
          > involved.
          >
          > "Namespace" - I suggest making this an option that is user tunable, but pick
          > a good default value that hopefully won't need to be tweaked all the time.
          >
          It would vaguely resemble ``StateDir''. Each separate StateDir signifies
          a unique application.
          Thus far, using StateDir in shmfs/tmpfs I'd use
          /dev/shm/apache/website1 and /dev/shm/apache/website2 for two separate
          applications. Something similar would apply to using a single memcached
          for multiple applications, no ?
          Of course, the default StateDir ``.state'' would mix things up a bit.
          > "Lock/Unlock" - Each individual operation of memcached is atomic.. if two
          > processes attempt to write to one location at the same time, they will be
          > serialized and one will not corrupt the other. The problem is you don't
          > necessarily know which one will actually win. This means you can't guarantee
          > state if you don't do locking to ensure the right one wins.
          Well, the original point of Lock and UnLock was to avoid corrupting the
          on-disk files, but your point is valid..
          > Consider an
          > incrementing counter. Assuming you do a get to read the value, increment it
          > by one, and then do a set to save it; each individual operation is safe
          > (get, set), but not the combination (get+set). If two hit at the same time,
          > the number would increment by 1 and not by 2 (one for each). I'd worry that
          > making lock and unlock a no-op would create new problems, especially as site
          > volume increases and chances of simultaneous updates increase.
          >
          Each user gets his own key (session-id in Apache::ASP) to write, apart
          from the generic ``application''.
          The case you're describing isn't handled in Apache::ASP already and it's
          probably because the developer (end user in our case) should concern
          himself with that.
          But we can work on that one too if you feel so inclined.
          > "No list of keys" - Usually I know what keys I'm stuffing into memcacheb and
          > I don't need to walk through the keys, or if I do I have a place outside of
          > memcache that has the keys to lookup. This one could get tricky. How would
          > you keep the list of keys in a second entry intact, especially if two
          > processes wanted to add a key at the same time?
          >
          According to perltie, tying hashes pretty much expects ``FIRSTKEY'' and
          ``NEXTKEY'' to work. But this is an inadequacy of Cache::Memcached::Tie
          which we may or may not address.
          I was (indirectly) asking Josh if these are actually needed for
          Apache::ASP Sessions to work, seeing that they're defined in
          Apache::ASP::Session but never actually called - not from within
          Apache::ASP at least. But someone is bound to have used / wants to use
          keys(%$Session) already..
          > Other thoughts/suggestions:
          >
          > Having not looked at your code or design this may not be applicable, but
          > consider making this generic, something that memcache or another cache
          > engine could be plugged into. If you're interested, and my time permits, I'd
          > be interested in working on part of this with you.
          >
          The thing is, Apache::ASP::State is too tightly bound to on-disk dbm
          files and would require a major rewrite to facilitate other storage
          engines. I suspect that's what held Josh from implementing
          Apache::Session storage. Come to think of it, there already is an
          Apache::Session::Memcached thing out there, so perhaps we should focus
          on making Apache::ASP::State work with an Apache::Session and friends
          back-end instead of hacking around.
          I'd feel more comfortable if we had help from the original author of
          Apache::ASP for this (or at least his blessing ;)
          Let's think this through the weekend and decide on Monday.

          Best Regards,
          Thanos Chatziathanassiou



          ---------------------------------------------------------------------
          To unsubscribe, e-mail: asp-unsubscribe@...
          For additional commands, e-mail: asp-help@...
        • Thanos Chatziathanassiou
          I had my svn repository disk die on me recently, but still have my working copy around and got some free time to hack it. It turned into a real Apache::Session
          Message 4 of 5 , Dec 17, 2009
          • 0 Attachment
            I had my svn repository disk die on me recently, but still have my
            working copy around and got some free time to hack it.
            It turned into a real Apache::Session session store back-end for
            Apache::ASP and seems to be working fine so far, although it is a bit
            rough around the edges.

            Due to ``internal'' and ``application'' not being valid keys for most
            Apache::Session implementations, I had them hard-coded to
            ``00000000000000000000000000000000'' and
            ``ffffffffffffffffffffffffffffffff'' respectively. The ugly part is
            Apache::ASP::State::ApacheSession::ServerAutoKeys, which creates these
            automatically if they're needed when a request arrives.
            I'm afraid it is prone to deadlock if there are multiple simultaneous
            first requests and the session store is something heavier than memcached
            (memcached and redis seem to be fine with it by design of course). Due
            to this, it is off by default, but can be turned on in the server
            configuration.
            The hard-coded 1MB limit item size in memcached is somewhat problematic
            for internal and I had it run out without compression by creating about
            28000 sessions. The space each session occupies in internal is fixed,
            regardless of its contents, right ?
            I suppose having the session manager clean up on a tighter schedule is
            an option with a fast backend though.

            Some benchmarks look promising:
            memcached running on localhost
            $ perl multi_http.pl --requests=1000 --concurrency=5
            --url=http://127.0.0.1:8080/asp/benchmarks/memcached/index.asp
            Child 001: 200 requests success: 200, failure: 0 in 1.560713 sec
            Child 003: 200 requests success: 200, failure: 0 in 1.564235 sec
            Child 004: 200 requests success: 200, failure: 0 in 1.577891 sec
            Child 002: 200 requests success: 200, failure: 0 in 1.630787 sec
            Child 005: 200 requests success: 200, failure: 0 in 1.565008 sec
            Parent total time: 1.66203 sec

            classic sessions running on shmfs/tmpfs via MLDBM::Sync::SDBM_File
            $ perl multi_http.pl --requests=1000 --concurrency=5
            --url=http://127.0.0.1:8080/asp/benchmarks/classic/index.asp
            Child 001: 200 requests success: 200, failure: 0 in 1.849258 sec
            Child 005: 200 requests success: 200, failure: 0 in 1.973691 sec
            Child 004: 200 requests success: 200, failure: 0 in 2.004447 sec
            Child 002: 200 requests success: 200, failure: 0 in 2.064989 sec
            Child 003: 200 requests success: 200, failure: 0 in 2.094894 sec
            Parent total time: 2.106399 sec

            I did not yet try with MLDBM::Sync::SDBM_File on an NFS mounted
            filesystem (the server that would be hosting that died along with my svn
            repository), but these results look tempting.
            Since I'm afraid my disk might also die on me and I'd have to redo
            everything from the start, I'm attaching what I have so far.

            Once again, this has not been (properly) tested. This quite ugly. It is
            most definitely NOT production-stuff material. It is not even ready.
            Caveat emptor. Still, it shows Apache::Session back-end can be made to work.


            Configuration for the file served above:
            for memcached version:
            ---8<---
            PerlSetVar AllowSessionState 1
            PerlSetVar AllowApplicationState 1
            PerlSetVar StateDB Apache::Session::Memcached
            #Memcached
            PerlSetVar ApacheSessionParams Servers
            PerlAddVar ApacheSessionParams 127.0.0.1:11211
            PerlAddVar ApacheSessionParams Namespace
            PerlAddVar ApacheSessionParams testing
            PerlAddVar ApacheSessionParams CompressThreshold
            PerlAddVar ApacheSessionParams 10000
            PerlAddVar ApacheSessionParams AutoCreateApplicationInternalKeys
            PerlAddVar ApacheSessionParams 1
            ---8<---
            (The author of Apache::Session::Store::Memcached has not implemented
            ``Namespace'' and since I was already playing with his module, I cheated
            and changed it so that it uses Cache::Memcached::Fast instead of plain
            Cache::Memcached)

            for classic version:
            ---8<---
            PerlSetVar AllowSessionState 1
            PerlSetVar AllowApplicationState 1
            PerlSetVar StateDB MLDBM::Sync::SDBM_File
            PerlSetVar StateDir /dev/shm/apache
            ---8<---

            Other tested back-ends:
            #Sybase
            PerlSetVar StateDB Apache::Session::Sybase
            PerlSetVar ApacheSessionParams DataSource
            PerlAddVar ApacheSessionParams
            dbi:Sybase:database=session_db;server=dev_server
            PerlAddVar ApacheSessionParams UserName
            PerlAddVar ApacheSessionParams username
            PerlAddVar ApacheSessionParams Password
            PerlAddVar ApacheSessionParams password
            PerlAddVar ApacheSessionParams Commit
            PerlAddVar ApacheSessionParams 1

            #SQLite3
            PerlSetVar StateDB Apache::Session::SQLite3
            PerlSetVar ApacheSessionParams DataSource
            PerlAddVar ApacheSessionParams dbi:SQLite:dbname=/tmp/session.db

            #Postgres
            PerlSetVar StateDB Apache::Session::Postgres
            PerlSetVar ApacheSessionParams DataSource
            PerlAddVar ApacheSessionParams DBI:Pg:dbname=dev_server;host=192.168.100.46
            PerlAddVar ApacheSessionParams UserName
            PerlAddVar ApacheSessionParams username
            PerlAddVar ApacheSessionParams Password
            PerlAddVar ApacheSessionParams password
            PerlAddVar ApacheSessionParams Commit
            PerlAddVar ApacheSessionParams 1

            #Redis via self-made Apache::Session::Redis using Redis-0.08
            PerlSetVar StateDB Apache::Session::Redis
            PerlSetVar ApacheSessionParams Server
            PerlAddVar ApacheSessionParams 127.0.0.1:6379
            PerlAddVar ApacheSessionParams Namespace
            PerlAddVar ApacheSessionParams testing
          Your message has been successfully submitted and would be delivered to recipients shortly.