I have indices of 800M and told MySQL to use 1500M of RAM. After starting MySQL it uses 1000M on Windows 7 x64.
I want to execute this query:
SELECT oo.* FROM table o
LEFT JOIN table oo ON (oo.order = o.order AND oo.type="SHIPPED")
WHERE o.type="ORDERED" and oo.type IS NULL
This finds all items not yet shipped . The execution plan tells me this:
My indices are:
type_order: Multiple index with type and order
order_type with order as first index value, followed by type
So MySQL should use the index type_order from RAM and then pick out the few entries with the order_type index. I'm expecting only about 1000 non shipped items, so this query should be really fast, but it isn't. Disks are going crazy....
What am I doing wrong?
The query says SELECT sometable.*, so for 1000 matching rows, there will be 1000 fetches of all the fields from the table. Whether the WHERE part indexes are fully loaded into ram or not would only help some. The data fields still have to be retrieved. Odds are, they are scattered all over the disk. So, of course the disk(s) will be doing a thousand small reads.
Related
I have created a 2TB MySQL RDS, and filled it with 2 tables totaling 1.5TB:
+----------+---------------------------+------------+
| Database | Table | Size in MB |
+----------+---------------------------+------------+
| stam_db | owl | 1182043.00 |
| stam_db | owl_owners | 393695.00 |
The instance was set with db.m6g.2xlarge size and 6000 provisioned IOPS.
I ran this query to return the first 10 rows (they are all distinct, no duplicated rows):
SELECT DISTINCT *
FROM owl
ORDER BY
name
LIMIT 10;
To my surprise, this query has been running for the last 2 hours...
Even more surprising, the "Free Storage Space" AWS metric started to decrease at a rate of 2.2GB/minute:
For some reason, Write IOPS suddenly risen to 600-700 per second:
READ IOPS went even higher, to about 1850 per second:
This brings total IOPS to around 2400-2500:
CPU Utilization remained in the low single digits:
I have a few questions:
Why would a SELECT DISTINCT statement cause such massive writes into the database?
Why would the SELECT DISTINCT try to read the entire DB, instead of just the first 10 rows?
Why isn't RDS using the 6000 allocated IOPS? The total IOPS are only about 40% of the allocated amount.
For future reference, here are the answers:
Q2) I think I found an explanation at https://www.percona.com/blog/2019/07/17/mysql-disk-space-exhaustion-for-implicit-temporary-tables/ -" The queries that require a sorting stage most of the time need to rely on a temporary table. For example, when you use GROUP BY, ORDER BY or DISTINCT. Such queries are executed in two stages: the first is to gather the data and put them into a temporary table, the second is to execute the sorting on the temporary table." So even regular SELECT with ORDER BY needs to re-read then whole table
Q1) The massive writes are caused by the temporary table created for the query, they can reach 100% of the original table.
Q3) Looks like MySQL code creating the temporary tables simply isn't efficient enough to utilize the entire 6000 IOPS
Try to use EXPLAIN to analyze your SELECT DISTINCT query. I bet it will include "Using temporary" and/or "Using filesort". With a large enough result set, these queries will use temporary disk space. But the more frequently you run these queries, the more disk space it uses.
I don't know why you use SELECT DISTINCT * if the rows are already distinct. This may cause the use of a temporary table unnecessarily.
Ideally your query should be:
SELECT *
FROM owl
ORDER BY
name
LIMIT 10;
Make sure there is an index on the name column, so it can skip the "Using filesort" by reading rows in the index order by name.
Why isn't it using the full provisioned IOPS? I would guess because MySQL is constrained by the code that builds temporary tables. It can't fill the temp tables fast enough to saturate a high number of IOPS. Perhaps if you were to run this query concurrently in many threads it would. But maybe not. IMO, provisioned IOPS are pretty much a scam.
SELECT t1.*
FROM
( SELECT key_a,key_b,MAX(date) as date
FROM large_table
WHERE date <= **20150126**
group by key_a,key_b
) AS t2
JOIN large_table AS t1 USING(key_a,key_b ,date)
large_table = 1,223,001,206 rows of data
Primary Key key_a,key_b,date
key on key_b
key on date
There are numerous empty dates between rows for a & b that I want the most recent behind or on the "Date" entered.
Is it the Mysql Join settings causing it to be slow ?
I can copy the entire set of a & b data with an INSERT to a temp table just by selecting all the rows and then run the same query on the temp table, but why do multi queries (insert selected, then select from) when only 1 is needed.
The query above only has 4,128,548 total results in the temp insert all dates table, and the date specific returns under 180,000 total.
Not table optimization, not keys, is it Max sort length, Join Buffer size , I have 128 gig ram, on a 32 core server running this, there is no reason for it to be slow, just never bulk insert this large of a single table to run Join queries on prior if anyone else has dealt with tables this size any info greatly appreciated.
Edited query, yes it's late long day had Distinct when it wasn't needed or in actual query
WHERE date <= **20150126**
group by key_a,key_b
needs an index starting with date. It's about doing what you can with the WHERE clause, not sparse or dense.
Then... Since the inner query references only 3 columns, building a 'covering' index may be useful. (Probably useful in your case.) So, tack on the other two fields, in either order. Such as
INDEX(`date`, key_a, key_b)
For MyISAM this step is critical. For InnoDB, this is redundant, since each secondary key (such as your INDEX(date)) implicitly includes the rest of the fields of the PK.
No, the PRIMARY KEY(key_a, key_b, date) cannot serve the purpose. It's in the wrong order. Also, it is (if you are using InnoDB) "clustered" with the index.
The query above only has 4,128,548 total results in the temp insert all dates table, and the date specific returns under 180,000 total.
Sorry, I had trouble parsing that. I assume you are saying 4M rows had 'date<...' and the subquery delivered only 180K rows. Hence, the outer query also returned 180K rows.
The first goal is to get through the 4M rows as efficiently as possible. With the index I propose, that might be about 20K blocks (#16KB each) of index scanning. That's 300MB.
Next the MAX and GROUP BY are performed. At 300MB, this will involve a disk tmp table. (See max_heap_size and max_tmp_table_size.)
Then comes the JOIN to fetch t1.*. You are using a good technique for fetching a bunch of rows from a huge table, where you need a GROUP BY (or LIMIT or ...) that is clumsy when done the obvious way. It goes like this: Write the subquery to find the PKs. Get the best index for it. Then JOIN on the PK.
Now for the JOIN. (Again, I assume InnoDB.) Since you are JOINing on the PK, each lookup into t1 will be efficient -- drill down the PK's BTree to find a row. Do that 180K times.
If those 180K lookups are scattered around the table, then this could be 180K disk hits.
Total effort: 20K + 180K = 200K disk hits, possibly less. On commodity spinning disks, this would take about 30 minutes (plus time for the tmp table). (No, only one core will be used. Anyway, I/O is probably the bottleneck.)
OPTIMIZE TABLE -- almost always useless.
I assume innodb_buffer_pool_size is about 90G? If things are going to be cached, that is where it would happen (for InnoDB). Since 200K blocks is 3GB, it could be easily cached. That is, if you run the query twice, the first might be 30 minutes, but the second might be less than 3 minutes.
To get more numbers, you could do:
FLUSH STATUS;
SELECT ...;
SHOW SESSION STATUS;
and look for 'Handler%', '%sort%', 'Innodb%' and maybe a few others.
What version are you running? Recent versions have a leapfrog technique that works better for max+groupby than what I described. I think it is called MRR. If so, your PK is actually optimal. (Hmmm... I should play around with that.)
PARTITIONing -- I don't see any benefit (for this query).
I am reading High performance MySQL and I am a little confused about deferred join.
The book says that the following operation cannot be optimized by index(sex, rating) because the high offset requires them to spend most of their time scanning a lot of data that they will then throw away.
mysql> SELECT <cols> FROM profiles WHERE sex='M' ORDER BY rating LIMIT 100000, 10;
While a deferred join helps minimize the amount of work MySQL must do gathering data that it will only throw away.
SELECT <cols> FROM profiles INNER JOIN (
SELECT <primary key cols> FROM profiles
WHERE x.sex='M' ORDER BY rating LIMIT 100000, 10
) AS x USING(<primary key cols>);
Why a deferred join will minimize the amount of gathered data.
The example you presented assumes that InnoDB is used. Let's say that the PRIMARY KEY is just id.
INDEX(sex, rating)
is a "secondary key". Every secondary key (in InnoDB) includes the PK implicitly, so it is really an ordered list of (sex, rating, id) values. To get to the "data" (<cols>), it uses id to drill down the PK BTree (which contains the data, too) to find the record.
Fast Case: Hence,
SELECT id FROM profiles
WHERE x.sex='M' ORDER BY rating LIMIT 100000, 10
will do a "range scan" of 100010 'rows' in the index. This will be quite efficient for I/O, since all the information is consecutive, and nothing is wasted. (No, it is not smart enough to jump over 100000 rows; that would be quite messy, especially when you factor in the transaction_isolation_mode.) Those 100010 rows probably fit in about 1000 blocks of the index. Then it gets the 10 values of id.
With those 10 ids, it can do 10 joins ("NLJ" = "Nested Loop Join"). It is rather likely that the 10 rows are scattered around the table, possibly requiring 10 hits to the disk.
Let's "count the disk hits" (ignoring non-leaf nodes in the BTrees, which are likely to be cached anyway): 1000 + 10 = 1010. On ordinary disks, this might take 10 seconds.
Slow Case: Now let's look at the original query (SELECT <cols> FROM profiles WHERE sex='M' ORDER BY rating LIMIT 100000, 10;). Let's continue to assume INDEX(sex, rating) plus the implicit id on the end.
As before, it will index scan through the 100010 rows (est. 1000 disk hits). But as it goes, it is too dumb to do what was done above. It will reach over into the data to get the <cols>. This often (depending on caching) requires a random disk hit. This could be upwards of 100010 disk hits (if the table is huge and caching is not very useful).
Again, 100000 are tossed and 10 are delivered. Total 'cost': 100010 disk hits (worst case), which might take 17 minutes.
Keep in mind that there are 3 editions of High performance MySQL; they were written over the past 13 or so years. You are probably using a much newer version of MySQL than they covered. I do not happen to know if the optimizer has gotten any smarter in this area. These, if available to you, may give clues:
EXPLAIN FORMAT=JSON SELECT ...;
OPTIMIZER TRACE...
My favorite "Handler" trick for studying how things work may be helpful:
FLUSH STATUS;
SELECT ...
SHOW SESSION STATUS LIKE 'Handler%'.
You are likely to see numbers like 100000 and 10, or small multiples of such. But, keep in mind that a fast range scan of the index counts as 1 per row, and so does a slow random disk hit for a big set of <cols>.
Overview: To make this technique work, the subquery need a "covering" index, with the columns correctly ordered.
"Covering" means that (sex, rating, id) contains all the columns touched. (We are assuming that <cols> contains other columns, perhaps bulky ones that won't work in an INDEX.)
"Correct" ordering of the columns: The columns are in just the right order to get all the way through the query. (See also my cookbook.)
First come any WHERE columns compared with = to constants. (sex)
Then comes the entire ORDER BY, in order. (rating)
Finally it is 'covering'. (id)
From the description below from official (https://dev.mysql.com/doc/refman/5.7/en/limit-optimization.html):
If you combine LIMIT row_count with ORDER BY, MySQL stops sorting as soon as it has found the first row_count rows of the sorted result, rather than sorting the entire result. If ordering is done by using an index, this is very fast. If a filesort must be done, all rows that match the query without the LIMIT clause are selected, and most or all of them are sorted, before the first row_count are found. After the initial rows have been found, MySQL does not sort any remainder of the result set.
We can see that they should have no difference.
But the percona suggest this, and give test data. But give no reason, I think there maybe exist some "bug" in mysql when deal with this kind of case. So we just regard this as a useful experience.
I'm using PHP & MySQL for my website and have to aggregate some metrics to generate a report for the client. The metrics can be found in a 4-million row table, which uses the MyISAM engine and is distributed in 12 PARTITIONS.
In a PHP loop (a for-loop), for each iteration, I retrieve 1000 rows that match specific ids with
id = X OR id =Y OR id = Z
(I'm not using any inner join with a temporary table like UNION 1 id UNION 2 etc. as it is a little bit slower, might be because of the partition option that relies on the hash of the id).
The problem is that the queries are getting slower and slower. It might be caused by something that is cached progressively, but I don't know what.
Any help would be very precious, many thanks.
MySQL gets very slow when you use a LIMIT that starts deep in the index. Check out this article for optimizing late row lookups:
http://explainextended.com/2009/10/23/mysql-order-by-limit-performance-late-row-lookups/
We're doing an update query between two database tables and it is ridiculously slow. As in: it would take 30 days to perform the query.
One table, lab.list, contains about 940,000 records, the other, mind.list about 3,700,000 (3.7 million)
The update sets a field when two BETWEEN conditions are met. This is the query:
UPDATE lab.list L , mind.list M SET L.locId = M.locId WHERE L.longip BETWEEN M.startIpNum AND M.endIpNum AND L.date BETWEEN "20100301" AND "20100401" AND L.locId = 0
As it is now, the query is performing with about 1 update every 8 seconds.
We also tried it with the mind.list table in the same database, but that doesn't matter for the query time.
UPDATE lab.list L, lab.mind M SET L.locId = M.locId WHERE longip BETWEEN M.startIpNum AND M.endIpNum AND date BETWEEN "20100301" AND "20100401" AND L.locId = 0;
Is there a way to speed up this query? Basically IMHO it should make two subsets of the databases:
mind.list.longip BETWEEN M.startIpNum AND M.endIpNum
lab.list.date BETWEEN "20100301" AND "20100401"
and then update the values for these subsets. Somewhere along the line I think I made a mistake, but where? Maybe there is a faster query possible?
We tried log_slow_queries, but that shows that it is indeed examining 100s of millions of rows, probably going up all the way to 3331 gigarows.
Tech info:
Server version: 5.5.22-0ubuntu1-log (Ubuntu)
lab.list has indexes on locId, longip, date
lab.mind has indexes on locId, startIpNum AND M.endIpNum
hardware: 2x xeon 3.4 GHz, 4GB RAM, 128 GB SSD (so that should not be a problem!)
I would first of all try to index mind on startIpNum, endIpNum, locId in this order. locId is not used in SELECTing from mind, even if it is used for the update.
For the same reason I'd index lab on locId, date and longip (which isn't used in the first chunking, which should run on date) this order.
Then what kind of datatype is assigned to startIpNum and endIpNum? For IPv4, it's best to convert to INTEGER and use INET_ATON and INET_NTOA for user I/O. I assume you already did this.
To run the update, you might try to segment the M database using temporary tables. That is:
* select all records of lab in the given range of dates with locId = 0 into a temporary table TABLE1.
* run an analysis on TABLE1 grouping IP addresses by their first N bits (using AND with a suitable mask: 0x80000000, 0xC0000000, ... 0xF8000000... and so on, until you find that you have divided into a "suitable" number of IP "families". These will, by and large, match with startIpNum (but that's not strictly necessary).
* say that you have divided in 1000 families of IP.
* For each family:
* select those IPs from TABLE1 to TABLE3.
* select the IPs matching that family from mind to TABLE2.
* run the update of the matching records between TABLE3 and TABLE2. This should take place in about one hundred thousandth of the time of the big query.
* copy-update TABLE3 into lab, discard TABLE3 and TABLE2.
* Repeat with next "family".
It is not really ideal, but if the slightly improved indexing does not help, I really don't see all that many options.
In the end, the query was too big or cumbersome for mysql to fill. Even after indexing. Testing the same query with the same data on a high-end Sybase server, also took 3 hours.
So we abandoned the do it all on the database server thought, and went back to scripting languages.
We did the following in python:
load a chunk of 100000 records of the 3.7 million records, and loop over the rows
for each row, set the locId and fill in the rest of the columns
All these updates together take about 5 minutes, so a huge improvement!
Conclusion:
think outside of the database box!