I wish he would do an explain on the original slow query. 12 million rows is not that much, this looks a lot like he was just missing an index on device_id.
What SQL was running, and what was the query plan?
ActiveRecord is notorious for generating terrible, terrible SQL.
Edit: What I would expect to see would be index scan on device id, then sort + limit. So the important factor would be rows per device, not total rows.
Queries like that should be very, very fast even with low amounts of ram. Good article, but consider using EXPLAIN a few times first before embarking on such adventures. :)
Definitely want to second the advice to use EXPLAIN.
MySQL has a lot of very specific limitations about when it will and will not use the available indices. It also matters how you created the indices (one index on multiple columns vs multiple indices on individual columns).
For example, if you created a single index with the columns (date_added, device_id) MySQL would not be able to use the index since device_id is the second part of the index, not the first and thus not available for use in the WHERE clause.
if you had an index on device_id there shouldn't be a difference in performance between the original table and the partioned one. the index is used to reduce the rows that need to be scanned so there wouldn't be a full table scan. btw, why do you use a table lock?
Were those single indexes like
CREATE INDEX index_1 ON Locations(id)
CREATE INDEX index_2 ON Locations(user_id)
...?
If you have multiple single column indexes it won't help you here. The mysql query planner probably correctly guessed the table scan was a better option than hash joining the results of two seeks given the machine's limited memory. It also probably underestimated the cost of disk IO on a virtual machine though.
MySQL can only use one index per query. If you need it be on two or more columns you need a composite index with both columns in, populated from the left. If MySQL finds a column in the index you not keying off (WHERE/SORT etc.) then it will stop processing the index at that point.
The key_len column in the EXPLAIN shows you how far it got through the index, based on the size of the columns. In your case it says 4 which if device id is of type INT it just read one column in the index (though there is probably only one column in that index)
While the write up was good, the author makes assumptions about what's happening seemingly without checking. He didn't post the actual query or the original and final query plans.
I can't know for sure since there's not enough information there, but I'd suspect a simple index would have done the trick.
CREATE INDEX my_index ON Location (device_id, date_added desc)
SELECT *
FROM Locations USE INDEX (my_index)
WHERE device_id = ...
ORDER BY date_added DESC
LIMIT 6
SELECT *
FROM Locations USE INDEX (my_index)
WHERE device_id = ...
AND date_added BETWEEN ... AND ...
ORDER BY date_added DESC
I'm assuming that he had the index and it just wasn't getting used, so the index hints are there in case the query planner was somehow missing them.
That said, table partitioning is an awesome feature. I've seen it be most useful when you spread the partitions over tiered storage (latest on SSD, archive on spindles), or when you want to drop a whole range of data quickly on a regular basis (like last month's logs).
That's entirely right; to explain a bit more, the reason that index is good is because it's a covering index. You have all the data you need to pull from the table in the index (which you've found via binary search through your B+tree), so you don't need to run out again to the table, and that saves an extra disk seek.
To make a covering index takes longer, so insertion speed drops, but that shouldn't be the hotspot.
Actually, since he was effectively doing SELECT * ... it still has to go pull the rows from the table after seeking the index. This likely isn't a huge problem though.
To prevent going back to the table in this case, you'd either need the index I mentioned to be the primary key, or a construct like INCLUDE (which mysql appears to lack).
Honestly, the first thing I did was add indexes. On device_id, user_id, and date_added.
I'm using Linode and I kept getting these I/O notifications...and the site would load extremely slow at times.
After looking at the log of my Rails app, I saw that some requests were taking > 20 seconds. I determined that this was some kind of blocking at the database.
I connected to MySQL and ran something simple...
select count(*) from location;
It took a long time...I turned profiling on and saw that it was taking a really long time in "table lock".
I assumed (which probably wasn't a good idea) that it was going through and counting all the rows.
I didn't really know what to make of this..because I didn't think that getting the count would take so long.
I embarked on reading about partitioning, which may have been a solution for a problem that didn't actually exist (based on the feedback here). I attempted to partition (on what I put in the article), and everything seemed much snappier after that. If the indexes should have solved the problem (given that they were correct), I don't know why the location queries were taking so long.
Anyway, I still have lots to learn on the database front, and maybe the fact that my VM had 7MB free of memory was causing weird things to happen, I'm not sure.
Thanks for all the feedback and I have definitely learned a lot in this thread.
kogir:
I'm using MyISAM (which, in retrospect, seems stupid), so I don't even have foreign key constraints (InnoDB only, I believe).
MySQL does not scan any rows for SELECT COUNT(*) FROM table — the number of records is stored in table metadata. (Mind you, it only works for this kind of query, as soon as you add WHERE the story changes).
On the other hand, MyISAM only has table level locks, not row level.
I guess it is time for you to analyze what your app does with DB and how it can affect locking…
Sort of true -- for the InnoDB engine SELECT COUNT(*) FROM table is more complicated because each transaction has its own view of the table. So unlike MyISAM (which uses one big table-level lock to guarantee consistency), it can't store a row count on a per-table basis.
1. Were the keys composite or not? The correct key for the queries you listed is (device_id, date_added DESC). In MySQL you'd add this by doing "ALTER TABLE location ADD INDEX (device_id, date_added DESC)"
MySQL uses one key per query, so having an index on (device_id) and an index on (date_added) individually will do no (or very little) good for the queries you listed.
2. MyISAM uses table-level locking. This will block all reads (selects) while writes (insert, update, delete) are happening. It does not prevent multiple selects from happening at the same time. If you have a write-heavy table that you are also reading from you need to be using InnoDB.
3. SELECT COUNT(*) FROM location should be instantaneous in MyISAM . It will be slow(er) in InnoDB. If this query was blocking it's because something was trying to write to the table.
4. Foreign key constraints have nothing to do with anything here. That's one advantage of InnoDB, although most web apps I know don't use them, but not one that would help you. Rather, what will help you is the table-level vs. row-level for MyISAM vs. InnoDB.
This allows selects to happen to unaffected rows while writes are happening.
22 comments
[ 2.7 ms ] story [ 48.4 ms ] threadActiveRecord is notorious for generating terrible, terrible SQL.
Edit: What I would expect to see would be index scan on device id, then sort + limit. So the important factor would be rows per device, not total rows.
The SQL was pretty simple I believe..select * from location where device_id = ? order by date_added desc limit 6...something like that.
Edit: Also, I don't know how much it matters, but MySQL probably only has ~256mb memory available to it (its hosted on a Xen box).
MySQL has a lot of very specific limitations about when it will and will not use the available indices. It also matters how you created the indices (one index on multiple columns vs multiple indices on individual columns).
For example, if you created a single index with the columns (date_added, device_id) MySQL would not be able to use the index since device_id is the second part of the index, not the first and thus not available for use in the WHERE clause.
See http://dev.mysql.com/doc/refman/5.1/en/mysql-indexes.html for more limitations.
If you have multiple single column indexes it won't help you here. The mysql query planner probably correctly guessed the table scan was a better option than hash joining the results of two seeks given the machine's limited memory. It also probably underestimated the cost of disk IO on a virtual machine though.
The key_len column in the EXPLAIN shows you how far it got through the index, based on the size of the columns. In your case it says 4 which if device id is of type INT it just read one column in the index (though there is probably only one column in that index)
Partitioning is a rather complex solution to his problem. See the limitations at http://dev.mysql.com/doc/refman/5.1/en/partitioning-limitati...
I can't know for sure since there's not enough information there, but I'd suspect a simple index would have done the trick.
I'm assuming that he had the index and it just wasn't getting used, so the index hints are there in case the query planner was somehow missing them.That said, table partitioning is an awesome feature. I've seen it be most useful when you spread the partitions over tiered storage (latest on SSD, archive on spindles), or when you want to drop a whole range of data quickly on a regular basis (like last month's logs).
* Edit: Missed something:
You may need to specify more hints than I originally thought if you want to force index usage for the order by clause as well. See http://dev.mysql.com/doc/refman/5.1/en/index-hints.html
To make a covering index takes longer, so insertion speed drops, but that shouldn't be the hotspot.
To prevent going back to the table in this case, you'd either need the index I mentioned to be the primary key, or a construct like INCLUDE (which mysql appears to lack).
See http://msdn.microsoft.com/en-us/library/ms188783.aspx
I'm using Linode and I kept getting these I/O notifications...and the site would load extremely slow at times.
After looking at the log of my Rails app, I saw that some requests were taking > 20 seconds. I determined that this was some kind of blocking at the database.
I connected to MySQL and ran something simple...
select count(*) from location;
It took a long time...I turned profiling on and saw that it was taking a really long time in "table lock".
I assumed (which probably wasn't a good idea) that it was going through and counting all the rows.
I didn't really know what to make of this..because I didn't think that getting the count would take so long.
I embarked on reading about partitioning, which may have been a solution for a problem that didn't actually exist (based on the feedback here). I attempted to partition (on what I put in the article), and everything seemed much snappier after that. If the indexes should have solved the problem (given that they were correct), I don't know why the location queries were taking so long.
Anyway, I still have lots to learn on the database front, and maybe the fact that my VM had 7MB free of memory was causing weird things to happen, I'm not sure.
Thanks for all the feedback and I have definitely learned a lot in this thread.
kogir:
I'm using MyISAM (which, in retrospect, seems stupid), so I don't even have foreign key constraints (InnoDB only, I believe).
query_cache_limit = 1M query_cache_size = 16M key_buffer = 16M max_allowed_packet = 16M thread_stack = 192K thread_cache_size = 8
1) Have you run Mysqltuner? If not, wget http://mysqltuner.com/mysqltuner.pl chmod to executable and execute it. I find it rather helpful.
2) Are you indexing every field that is part of any WHERE statement (at least for the big tables)?
3) Have you run EXPLAIN on the original query? (Suggested by someone else in this thread, and a critical question.)
4) Is the slow_query_log enabled, and are you sure about which query is causing the slowdown?
5) Have you tried tools like MySQL Administrator, which lets you monitor connections, threads, keys, queries, etc, in realtime?
MySQL uses one key per query, so having an index on (device_id) and an index on (date_added) individually will do no (or very little) good for the queries you listed.
2. MyISAM uses table-level locking. This will block all reads (selects) while writes (insert, update, delete) are happening. It does not prevent multiple selects from happening at the same time. If you have a write-heavy table that you are also reading from you need to be using InnoDB.
3. SELECT COUNT(*) FROM location should be instantaneous in MyISAM . It will be slow(er) in InnoDB. If this query was blocking it's because something was trying to write to the table.
4. Foreign key constraints have nothing to do with anything here. That's one advantage of InnoDB, although most web apps I know don't use them, but not one that would help you. Rather, what will help you is the table-level vs. row-level for MyISAM vs. InnoDB.
This allows selects to happen to unaffected rows while writes are happening.