We've also got our own pool implementation at OpenX
https://github.com/openx/gen_server_pool
The nice thing is it wraps gen_server such that you don't actually
have to know there is a pool there at all. Its not well documented
and currently uses mondemand to send stats (but I'll be trying to
remove that as a hard dependency). We use it to pool riak connections
and thrift connections and it works great.
An example of it's use for riakc would be
-module (my_riak_client).
start_link () ->
gen_server_pool:start_link (
{local, my_pool}, % identifier of pool
riakc_pb_socket, % gen_server to start
["localhost",8087,[{auto_reconnect, true}]], % options for init/1
[], % I don't recall
[ { prog_id, my_riak_client}, % mondemand identifier
{ min_pool_size, 1 }, % minimum 1 connection
{ max_pool_size, 10 }, % maximum 10 connections
{ idle_timeout, 60 }, % after 60 seconds of idle kill connections
{ pool_id, my_pool } ] % same identifier as first argument
).
% then when you use it
riak_get (Bucket, Key) ->
riakc_pb_socket:get (my_pool, % pool identifier
Bucket,
Key).
Behind the scenes its doing checkin/checkout and whatnot. It does look
cheat a bit by using calling the embedded gen_server functions directly
which actually caused us some consternation recently as we did account for
gen_server:reply/2 which riakc_pb_socket uses. However, we patched it and
now it works for those cases as well.
-Anthony
On Mon, Sep 24, 2012 at 01:20:33PM +0400, Dmitry Demeshchuk wrote:
> My main concern about poolboy is the fact that it wraps another gen_server
> process around the actual worker process. This makes you send an extra
> message every time you query the database, which is quite an overhead when
> you have a lot of concurrent users. Since you guys are fighting for
> performance, that could be an issue for you.
>
> At Mochi, we have a client of our own, which is somewhat close to gen_pool
> behaviour I implemented and open-sourced. This client incomplete in many
> ways, and uses parametrized modules, which a lot of people tend to dislike,
> but it is very fast, stable and production-proven.
>
> Nevertheless, I would really like to hear about Basho's opinion on poolboy
> too. Not only they included it into Riak distribution, but they also spent
> a handful of effort to test it through and fix a number of quite serious
> errors.
>
> On Mon, Sep 24, 2012 at 12:56 PM, Yuri Lukyanov <[email protected]> wrote:
>
> > Hi Dmitry, ;)
> >
> > Yes, thank you. I'm also thinking about this approach. I just wanted
> > to know how people use poolboy for that (It seems they do).
> > But even considering your case, imagine that you have quite a big
> > cluster and quite a big amount of rps.
> > In this case you may want to have more than one load balancer to
> > spread the load and be more tolerant to LB failures.
> > Exactly the same situation arises.
> > It's not a problem to implement a pooler in this case on my own. I'm
> > just asking about using poolboy for that.
> >
> > On Mon, Sep 24, 2012 at 12:08 PM, Dmitry Demeshchuk
> > <[email protected]> wrote:
> > > Why not use load balancer on top of Riak cluster, independently from
> > > clients? If load balancing is sophisticated enough, it can even do a
> > greater
> > > job than just uniformly spreading requests between machines.
> > >
> > > Consider the following situation. You happen to send several requests
> > for a
> > > huge piece of data (each can be just a simple get, or even a complicated
> > > map-reduce query), but the rest of the nodes, or just some of them,
> > receive
> > > much smaller queries.
> > >
> > > What an external load balancer can do is gathering some basic stats from
> > > each machine (like load average, memory and IO consumption, number of
> > rps to
> > > Riak and so on) and balancing the requests according to this data.
> > Another
> > > benefit from this approach is that you can easily add any other Riak
> > client
> > > (Python, Ruby, whatever else) on top of it, and load will be still nicely
> > > distributed between the machines.
> > >
> > > On Mon, Sep 24, 2012 at 11:38 AM, Yuri Lukyanov <[email protected]>
> > wrote:
> > >>
> > >> Hi,
> > >>
> > >> I'm trying to use poolboy to organize a pool of connections to riak
> > >> nodes with protobuf client.
> > >>
> > >> This has come from the suggestion here:
> > >>
> > >>
> > http://lists.basho.com/pipermail/riak-users_lists.basho.com/2012-September/009346.html
> > >>
> > >> The quote:
> > >> "Using Poolboy is convenient because it comes as a dependency of
> > >> riak_core.
> > >>
> > >> If you use Poolboy, you'll have to modify riakc_pb_socket slightly to
> > >> account for the way poolboy initializes connections (add a
> > start_link/1),
> > >> or create a simple module to pass the initialization from poolboy to
> > >> riakc_pb_socket."
> > >>
> > >> What I'm pondering about is what would be a convenient way to start
> > >> riakc_pb_socket clients in a pool.
> > >> For example, if I have 10 riak nodes in a cluster and want to open a
> > >> single connection to each of them (using riakc_pb_socket), I need one
> > >> poolboy worker per one riakc_pb_socket instance. In this case, I need
> > >> a host/port of a single riak node for each
> > >> riakc_pb_socket:start_link(). So I use WorkerArgs in
> > >> poolboy:start_link/3 to pass a list of riak nodes. And here comes a
> > >> confusion. poolboy passes WorkerArgs to each of its workers, so each
> > >> worker gets the list of riak nodes. Now I need to somehow choose a
> > >> node, maybe like proposed here:
> > >> https://github.com/basho/riak-erlang-client/pull/45/files
> > >> But I don't think it's a good solution.
> > >>
> > >> Maybe my idea of using riakc_pb_socket together with poolboy is not
> > >> convenient? Maybe I should start a pool per each riak node? Maybe
> > >> poolboy is not that convenient in this case as it seems?
> > >>
> > >> Would you share your experiences and thoughts on the subject?
> > >>
> > >> Thank you.
> > >>
> > >> _______________________________________________
> > >> riak-users mailing list
> > >> [email protected]
> > >> http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
> > >
> > >
> > >
> > >
> > > --
> > > Best regards,
> > > Dmitry Demeshchuk
> >
>
>
>
> --
> Best regards,
> Dmitry Demeshchuk
> _______________________________________________
> riak-users mailing list
> [email protected]
> http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
--
------------------------------------------------------------------------
Anthony Molinaro <[email protected]>
_______________________________________________
riak-users mailing list
[email protected]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
