Thread: Indexing queries with bit masks

Mike Christensen <mike@kitchenpc.com> writes:
> When a certain event happens, let's say event 4, I need to query for which
> users to notify.  So I'll be doing something like:

> SELECT UserId FROM Users WHERE Subscriptions & 8;

> My question is say there's a million rows in the Users table.  If I have an
> index on Subscriptions, will this index be used in the above query?

No.  At least not with a standard btree index.

I'm not exactly sure that an index would be helpful at all --- it seems
like the selectivity of this condition won't be very good anyway, will
it?  The more popular notifications will be subscribed to by a large
fraction of the user base.  Maybe it'd be useful to index unpopular
notifications, but how often will you be searching for those?

            regards, tom lane

On Fri, Apr 30, 2010 at 10:08 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Mike Christensen <mike@kitchenpc.com> writes:
>> When a certain event happens, let's say event 4, I need to query for which
>> users to notify.  So I'll be doing something like:
>
>> SELECT UserId FROM Users WHERE Subscriptions & 8;
>
>> My question is say there's a million rows in the Users table.  If I have an
>> index on Subscriptions, will this index be used in the above query?
>
> No.  At least not with a standard btree index.
>
> I'm not exactly sure that an index would be helpful at all --- it seems
> like the selectivity of this condition won't be very good anyway, will
> it?  The more popular notifications will be subscribed to by a large
> fraction of the user base.  Maybe it'd be useful to index unpopular
> notifications, but how often will you be searching for those?
>

We've got some similar columns (though nothing with any major number
of rows), so this is interesting...

If all subscriptions are roughly equal in popularity then any single
select should give ~ 10% of the data.  That would seem to be selective
enough that you'd really want an index?  If so, any answers to the
OP's main question; what would be the most efficient way to handle
this type of thing?

--
Peter Hunsberger

Peter Hunsberger <peter.hunsberger@gmail.com> writes:
> If all subscriptions are roughly equal in popularity then any single
> select should give ~ 10% of the data.  That would seem to be selective
> enough that you'd really want an index?

My personal rule of thumb is that 10% is around the threshold where
indexes stop being very helpful.  At that selectivity, you're going
to be having to read every page of the table anyway, and it's not
clear that the extra I/O to read the index is going to get repaid in
CPU savings.  (Now if the table+index are fully cached in RAM, the
threshold's probably a bit higher, but there still is not reason to
think that an index is going to make for a huge improvement.)

> If so, any answers to the OP's main question; what would be the most
> efficient way to handle this type of thing?

Well, btree's right out for indexing bit selections.  In principle you
could maybe do something with a GIN index, but I don't think we ship
any prefab GIN opclasses for this.

[ thinks for a bit ]

The best idea that comes to mind offhand is to not use an integer, but a
boolean array, such that the queries look like

    select ... where subscriptions[4];

This already gives you one big advantage, which is that you're not
hard-wiring an assumption about how many notification types there can
ever be.  What I would then do is build a separate partial index for
each subscription column, ie,

    create index ... where subscriptions[1];
    create index ... where subscriptions[2];
    .. etc ..

Now this only works as long as the queries are referencing explicit
constant subscription numbers, else the planner won't be able to
match the WHERE clause to any of the partial indexes.  But if that
is a reasonable restriction for your app then it seems like it
should work.

The main disadvantage of this is that you need N indexes, which could
get a bit expensive if the table is updated heavily.  But you don't need
to bother maintaining indexes corresponding to subscriptions that are
too popular to be worth indexing, so some of that could be bought back
by careful index selection.

Another point is that the partial indexes could be created on some other
column(s) and thereby serve double duty.  This depends on the details of
your typical queries though.  Is the subscriptions[] clause usually used
by itself, or together with additional WHERE conditions?

            regards, tom lane

On Fri, Apr 30, 2010 at 11:29 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Peter Hunsberger <peter.hunsberger@gmail.com> writes:
>> If all subscriptions are roughly equal in popularity then any single
>> select should give ~ 10% of the data.  That would seem to be selective
>> enough that you'd really want an index?
>
> My personal rule of thumb is that 10% is around the threshold where
> indexes stop being very helpful.  At that selectivity, you're going
> to be having to read every page of the table anyway, and it's not
> clear that the extra I/O to read the index is going to get repaid in
> CPU savings.  (Now if the table+index are fully cached in RAM, the
> threshold's probably a bit higher, but there still is not reason to
> think that an index is going to make for a huge improvement.)
>
>> If so, any answers to the OP's main question; what would be the most
>> efficient way to handle this type of thing?

Ok, that makes sense, which immediately makes me wonder if partitions
might make sense for this use case?  In particular if there really are
only 10 different types?

[...]

> The best idea that comes to mind offhand is to not use an integer, but a
> boolean array, such that the queries look like
>
>        select ... where subscriptions[4];
>

Interesting idea.  That might be worth testing for some of my use cases....

--
Peter Hunsberger

2010/4/30 Mike Christensen <mike@kitchenpc.com>:
> Ok I've been blatantly lying, err, purposely simplifying the problem for the
> sake of the original email :)
>
> I've read over the responses, and am actually now considering just not using
> any index at all.  Here's why:
>
> First, this actually isn't the only thing on the WHERE clause.  It will only
> query for users who are "friends" with you so it can notify them of your
> activities.  That's done via a weird JOIN on a table that holds all the
> friend relationships.  So in reality, it will only load maybe a hundred
> rows, or maybe a thousand every once in a while if you're way popular.  If
> I'm not mistaken, it should use the index to narrow it down to the list of
> friends, and then use a sequential scan to weed out the ones who subscribe
> to that type of notification.
>
> Second, the only thing /ever/ that will do this query is the queue service
> whose job it is to process notifications (which are files dropped on the
> file system) and email people all day long.  This service handles one job at
> a time, and could potentially run on its own machine with its own read-only
> copy of the database.  Thus, even if it was a fairly slow query, it's not
> gonna bring down the rest of the site.
>
> Regarding the idea of putting an index on each bit, I thought about this
> earlier as well as just kinda cringed.  The users table gets updated quite a
> bit (last logon, session id, any time they change their profile info,
> etc)..  Too many indexes is bad.  I could just put the data in another table
> of course, which lead me to another idea.  Have a table called Subscriptions
> and have each row hold a user id and a notification type.  I could index
> both, and join on (Subscriptions.UserId = Users.UserId AND
> Subscriptions.Type = 8).  This would be pretty dang fast, however updates
> are kinda a royal pain.  When the user changes which types of subscriptions
> they want (via a list of checkboxes), I'd have to figure out which rows to
> delete and which new ones to insert.  However, I think I have an idea in
> mind for a PgSQL function you pass in the bitmask to and then it
> "translates" it to conditional deletes and inserts.
>
> A third idea I'm tossing around is just not worry about it.  Put the bitmask
> in the DB, but not filter on it.  Every "friend" would be loaded into the
> dataset, but the queue processor would just "skip" rows if they didn't
> subscribe to that event.  In other words, move the problem down to the
> business layer.  The drawback is potentially large number of rows are
> loaded, serialized, etc into memory that will just be ignored.  But of
> course the DB is probably a read-only copy and it's not even close to the
> bottle neck of the email queue under heavy load, so it's probably a
> non-issue.  If mailing is slow, I just add more queue services..
>
> I'm exploring all these ideas.  I predict using the bitwise AND on the where
> clause isn't gonna be the worst design ever, and it's sure easier to
> implement than a table of subscriptions.  What do you guys think?

I would say "normalize". Which means that I like your "separate table"
idea best.
It's clear, obvious, and 3NF - conforming solution.
Changing the set of subscriptions with delete-update-insert combo is
not so bad as you would think.
Encapsulating it in some kind of functional API looks nice too.

Filip