Activerecord include queries - performance

rubyonrails-talk
1

Hi,

I noticed ActiveRecord uses some join queries to perform :include stuff. While this certainly fixes the N+1 problem, I'm wondering about the amount of sparse data this generates.

Example: Task has many TodoItems Both Task and TodoItem only have a name to keep things simple.

Now, when I do : Task.find(:all, :include => :todo_items)

it executes: SELECT tasks."id" AS t0_r0, tasks."name" AS t0_r1, todo_items."id" AS t1_r0, todo_items."name" AS t1_r1, todo_items."task_id" AS t1_r2 FROM tasks LEFT OUTER JOIN todo_items ON todo_items.task_id = tasks.id

which will transfer back: 1|task 1|1|item 1|1 1|task 1|2|item 2|1 1|task 1|3|item 4|1 1|task 1|4|item 3|1 2|task 2|5|item 1|2 2|task 2|6|item 2|2 2|task 2|7|item 3|2 2|task 2|8|item 4|2 3|task 3|9|item 1|3 3|task 3|10|item 2|3 3|task 3|11|item 3|3 3|task 3|12|item 4|3

As you can see, all fields for tasks are repeated for every todo item. While in this small example this wouldn't matter at all, my guess is that it _will_ matter more when there are more fields to a task, more items, or deeper nested joins. If I want to join data from say, 5 tables, all task data will get repeated a lot of times. While using mysql through a socket interface, I think it's neglectable (since queries that large will be quite slow anyway), but when connection to an external sql server, all this sparse data has to get transferred over the network.

Of course it's still a lot faster than not using the include, since looping over tasks would then generate: SELECT * FROM tasks SELECT * FROM todo_items WHERE (todo_items.task_id = 1) SELECT * FROM todo_items WHERE (todo_items.task_id = 2) SELECT * FROM todo_items WHERE (todo_items.task_id = 3)

(n+1) queries which is worse.

But how about this: tasks=Task.find(:all) todos = TodoItem.find(:all, :conditions => { :task_id => tasks.map {| x> x.id} })

what will generate: SELECT * FROM tasks SELECT * FROM todo_items WHERE (todo_items."task_id" IN (1,2,3))

which is 1 query per table. of course after this you'd have to find a way to 'link' the retrieved todos to the corresponding tasks, but I guess activerecord would be able to handle this (since it does it for the current, left-outer-join behaviour as well).

Now, I'm no database expert so I might be overlooking something. Also I haven't really done any benchmarks. But like I said, it looks to me as if the current ActiveRecord behaviour is quite sparse for transfering/parsing the data as soon as you 3 or 4 tables, with long rows, and many "children". Am I overlooking something?

Any thoughts on this matter would be nice.

Thanks, Mathijs

2

Hi,

I noticed ActiveRecord uses some join queries to perform :include stuff. While this certainly fixes the N+1 problem, I'm wondering about the amount of sparse data this generates.

It's not ideal at all (and compounded by some inefficiencies in the
way the data that is returned is parsed). There is a ticket/plan to
reimplement this as separate queries.

Fred

3

I noticed ActiveRecord uses some join queries to perform :include stuff. While this certainly fixes the N+1 problem, I'm wondering about the amount of sparse data this generates.

[...]

Now, I'm no database expert so I might be overlooking something. Also I haven't really done any benchmarks. But like I said, it looks to me as if the current ActiveRecord behaviour is quite sparse for transfering/parsing the data as soon as you 3 or 4 tables, with long rows, and many "children". Am I overlooking something?

Any thoughts on this matter would be nice.

You are unconsciously assuming that the cost of sending extra data is usually greater than the cost of associating records outside the database, not to mention multiple database optimize and execute cycles. A properly indexed database is going to perform the association of records very efficiently for you, since that's one of the things for which RDBMS software is designed and optimized. Note that YMMV with poorly written RDBMSs; I've had MySQL's query optimizer (not the execution engine, the query optimizer) completely hang when trying to deal with joins across a lot of tables (18, in this case).

The real cost of sending lots of data is, of course, more than just bandwidth. There is the cost of manifesting the result rows in the RDBMS and serializing them, which can cause thrashing if the result set is really big and have an even greater cost. There is the parsing cost for the database driver to bring the received data into memory. There is the cost of hydrating the models from the result set in memory.

These costs must be balanced against the smaller costs of the same thing done multiple times, including query overhead, to produce the multiple result sets, then performing the record association in memory. Incidentally, if you have the right indices for a WHERE IN clause to be efficient, you have the right indices for the joins to be efficient, so that's a wash.

The real question is where the line is such that the cost of dealing with the number of rows, M, produced by a multi-table join is greater than the cost of multiple queries and associating records in code. I think you need a whole lot of superfluous rows to win with multiple queries, but I could be wrong. This is where benchmarking comes in. Please do investigate.

Thanks, Mathijs

--Greg