Partitioning Presentation from MySQL Conference

It appears that my presentation isn’t available at the conference site. I’ve added it below.   Sorry for the delay. 

Without some context that I talked about at the conference the presentation may not make sense.

Partitioning creates a grain in the table.  Select queries that go with that grain can be quicker, at times much faster, but select queries that go against that grain can be slower, at times much slower. 

An example is a table that is partitioned by date and has an orderId primary key.  Queries that select data by date will either be almost as fast (partitioning can add a slight overhead) as the non-partitioned table to much faster.   But queries that don’t query by date will have to query all the partitions.  A table partitioned 12 times, one partition for each month, results in 12 index partitions.  As the indexes are partitioned by date, a query by the primary key of orderId means all twelve index partitions must be queried to look for the order in question.  Querying twelve index partitions is  going to be slower than just one unpartitioned table. 

Having said that, there are cases where partitioning dramatically improves performance, such as optimizing a table partition by partition. 

But now I’m starting to cover too much of what is in the presentation.  So here it is. 

http://www.slideshare.net/brooksaix/divide-and-be-conquered

Two New Databases Announced

 

The open source coderati announced a new database today called “Monsoon”.  Norm “Al” Modelle said much effort was spent by the team so it can handle internet sized floods of data.  He expected it to have the licensing of PostgreSQL, the multi-engine support of MySQL, the functionality of Oracle, the ease of use of MS SQL Server, and the massively parallel scalability of DB2.

Not to be outdone, Computer Dilettantes (CD) also announced a new database, the oddly named “Mud Puddle”. Eye Samme, the Executive VP in charge of databases we intend to milk for support revenue, said he knows the customer is being squeezed by licensing costs in these hard times.  So CD spent considerable effort designing a real time application that constantly analyzes the customer’s use of the database.  It then notifies the sales force how they can maximize revenue extraction from the customer. He said the database will have the licensing of Oracle, the ease of use of DB2, the single engine support of PostgreSQL, the scalability of MySQL, and the functionality of MS SQL Server.

Oh, and please attend MySQL partitioning session if you are thinking about implementing partitioning.  I promise it will have a bit more technical content. 

Is SQL Slow?

Last time I demonstrated a case where stored procedures are slow when they have to do computationally expensive work.  The more interesting, to me at least, question is how slow is SQL?  The answer to that is far more complex. 

For example, this simple SQL takes 2.2 seconds to run on MySQL.  This is a painfully faster than the 696 seconds it took a stored procedure to produce similar results. 

select customerName, count(*)
  from Salet s
  join customer c on s.customerId = c.customerId
group by customerName

As demonstrated in the previous article, the equivalent C# code took 1.4 seconds to produce the same results.  Some may find it surprising that it take less time to ship 1 million rows out of the database and then summarize it in code than it does for MySQL to summarize the same data in the database. 

In this simple case the performance difference isn’t much and is not worth the extra code complexity.  For more complex SQL, with more joins, perhaps nested case statements and temp tables, and similar standard SQL techniques, it is often much faster to do the transformation logic in non-SQL code.   I’ve found cases where I was able to improve performance by over ten times by using C# or Java code over SQL.  Still, my inclination is to always see if I can’t get acceptable performance from SQL first as less code is generally required, and usually far less code.  SQL is an exceptionally expressive language for certain types of problems. 

Plus, the performance advantage of C# or Java won’t be true in all cases.  Stating the obvious,  shipping out all the sales data won’t make sense for more selective queries that query only a few sales.  In this case it makes far more sense to write in SQL. 

Deciding where to execute code, in the database or elsewhere, is a complex problem that would require a series of articles to answer reasonably (a hint, think about using sequential IO as much as possible).  For now I just want to point out that running everything in SQL isn’t always the best performing method. 

Stored Procedures Are Slow Part 2

Last time I demonstrated a case where stored procedures are slow.  There were a few comments that I should include selects in the stored procedure to make the tests  more realistic.  From experience I already knew the answer so I didn’t go into that level of detail, but since stored procedures are all about database access this is a reasonable comment. 

This is a simple stored procedure that selects data and then summarizes it by customer.  No one would actually write this as it is far too easy to use a SQL statement instead.    Assume the logic is more complex and can’t be done easily by the standard SQL techniques of case statements, temp tables, etc, and then this makes more sense.  

The end result is this stored procedure takes 696 seconds to run. 

 

create procedure slowCounter()
begin

    declare done int default 0;
    declare curCustomerId int default 0;
    declare customerCount int default 0;
    declare slowCur cursor for select customerId from saleT;
    declare continue handler for not found
        set done = TRUE;

    create temporary table tempCount
    (
        tempCustomerId int not null,
        tempCustomerCount int not null,
        primary key (tempCustomerId)
    );

    open slowCur;
    fetch slowCur into curCustomerId ;
    while not done do
        insert into tempCount(tempCustomerId, tempCustomerCount)
            values (curCustomerId , 1)
        on duplicate key update tempCustomerCount = tempCustomerCount + 1;
        fetch slowCur into curCustomerId ;
    end while;

    close slowCur;

    select * from tempCount;

end

 

This more complex C# code (and still not fully exception proofed) only takes 1.4 seconds to run.  That is nearly 1/500 the time it took the stored procedure to execute.  

 

using System;
using System.Collections.Generic;
using MySql.Data.MySqlClient;

namespace Summarization
{
    /// <summary>
    /// Tests the performance of C# to simulate a SQL group by
    /// </summary>
    class SumTestGroup : BaseSum
    {
        protected MySqlConnection sqlConn;

        private Dictionary<Int32, String> customer =
            new Dictionary<Int32, String>();
        private Dictionary<String, Int32> customerRowCount =
            new Dictionary<String, Int32>();

        public void execute()
        {
            try
            {
                sqlConn = getConnectionMySql();
                sqlConn.Open();
                int count = 0;
                Int32 customerId;
                String customerName;

                getCustomers();

                MySqlCommand cmd = new MySqlCommand("select s.customerId from Salet s", sqlConn);
                MySqlDataReader rdr = cmd.ExecuteReader();

                while (rdr.Read())
                {
                    customerId = (Int32)rdr["customerId"];
                    customerName = customer[customerId];
                    count++;

                    if (customerRowCount.ContainsKey(customerName))
                    {
                        customerRowCount[customerName]++;
                    }
                    else
                    {
                        customerRowCount[customerName] = 1;
                    }
                    //Console.WriteLine("Count: " + count);
                }

                foreach (KeyValuePair<String, Int32> customerRow in customerRowCount)
                {
                    Console.WriteLine("Customer: " + customerRow.Key + " Customer Count: " +  customerRow.Value);
                }
                rdr.Close();

            }
            finally
            {
                sqlConn.Close();
            }
        }

        private void getCustomers()
        {
            MySqlCommand cmd = new MySqlCommand("select customerId, customerName from customer", sqlConn);
            MySqlDataReader rdr = cmd.ExecuteReader();

            Int32 customerId;
            String customerName;

            while (rdr.Read())
            {
                if (!rdr.IsDBNull(0))
                {
                    customerId = (Int32)rdr["customerId"];
                    customerName = (String)rdr["customerName"];
                    customer.Add(customerId, customerName);
                }
            }
            rdr.Close();
        }
    }
}

Both produce results like this:

customer Name 0, 100287
customer Name 1, 100009
customer Name 2, 100130
customer Name 3, 99655
customer Name 4, 100134
customer Name 5, 100102
customer Name 6, 99854
customer Name 7, 100172
customer Name 8, 99846
customer Name 9, 99812

 

In this case, the effort required to retrieve one million rows out of the database and then process them takes orders of magnitude less than a stored procedure takes to process the data inside the database.  

No one would actually write this code when a simple SQL alternative is possible.  But given complex logic that isn’t possible or easy to express in SQL that needs to be applied against a large volume of data, using a stored procedure cursor (or loop in general) isn’t a fast way to go.  This is odd as the whole point of stored procedures is to write code that deals with data contained in a database.  

Are MySQL stored procedures slow?

 

Yes, if compared to code in Java or C#. For example, this overly simple code took 284 seconds.

CREATE PROCEDURE CountTest()
begin
    declare counter int default 0;
    select now();
    repeat
        set counter = counter + 1;
    until counter > 120000000
    end repeat;
    select counter;
    select now();
end

Ignoring my off by one error, here is equivalent code in C# (the language I’m currently learning).  It took 419 milliseconds, or MySQL took 677 times longer to execute. From my experience, Java isn’t going to be any slower.

 int counter = 0;
while (counter < 120000000)
{
    counter++;
}

Slow stored procedure performance is one of the reasons why it usually isn’t wise to implement computationally expensive business logic in the database.   With networks not being the bottleneck they once were it is often better to extract the data to the application layer and process it in the programming language of your choice, even if that takes a few extra round trips to the database.  There are exceptions where those extra round trips are too numerous and prohibitively expensive.  With something as complex as a database there are always exceptions.

This example is a bit too simple, but, based on experience, it is representative of the general performance of stored procedure logic.

How can MySQL get away with such poor performance? Well, the equivalent code in MS SQL Server took 80 seconds on the same hardware, which is also two orders of magnitude slower than C# or Java code. 

create procedure testCount
as
begin
    declare @countNo int;
    set @countNo = 0;
    while (@countNo < 120000000)
    begin
        set @countNo = @countNo + 1
    end
    select @countNo
end

MS SQL Server code runs 3.5 times faster than MySQL, but given how much older MS SQL Server is, MySQL is doing well here.  So are MySQL Stored procedures slow?  Not really, if compared to other databases I’ve used.

Just don’t use them to do computationally expensive business logic.

Implementing Sharding in the Database

Over the past few weeks (years really) there has been some discussion on sharding. Instead of discussing when sharding is required, as there are good discussions on this already, I want to discuss how I would like to have sharding implemented in the database.

I want the database to handle sharding automatically, and where it can't be automatic, I want the database to help as much as it can.  Just like I want my business logic in a language ideally suited to it, and not stored procs (generally, there are always exceptions); I want all my physical persistence to be handled by the place that already does most of it, the database.  Having the database handle some of the physical persistence and the object relational layer handle the sharding logic isn’t ideal to me, and not just because the current object relational layers don’t have all the sharding functionality I want.  So here is what I want my database to do.  

1) I want to automatically route simple queries to the one shard that has the relevant data.  Hibernate shards already does this; I just want it implemented in the database.  The MPP (massively parallel processing) version of DB2 can also do this.

2) I need the ability to reshard data when a rebalancing of the shards is required, without downtime.  This is similar in concept to an index rebuild being done while the database is still up, which the expensive databases have implemented.  This is more complex than an index rebuild as multiple tables will need to be resharded before the resharding is complete. DB2 has functionality to reshard, but it requires down time last time I checked (which was a few years ago) .

3) Finally, and this is a superset of #1, I want to be able to run one SQL statement across multiple shards at the same time.  The database will send the SQL to all the relevant shards, run as much of the SQL on that shard as possible, using something like the function shipping that DB2 MPP has, and then aggregate the data from the various shards into one result set.  All I need to do this is execute one SQL statement.  Stated differently, all the shards will look like one database.  This won’t be easy to implement, and will probably be limited in functionality at first, but this long term goal is why implementing sharding in the database makes sense. 

This might sound a bit like an advertisement for the MPP version of DB2, but if it were open source I would at strongly consider implementing it for applications that need database sharding. Given that that is not likely to happen, I hope a reliable open source database, like MySQL, or Drizzle, or PostgreSQL, starts implementing this reasonably soon.

This sharding functionality, for the types of problems I’m seeing, is more important than any other new MySQL functionality, including better scaling on multi-core systems. With all above functionality, I can reasonably easily have multiple shards on one machine. Not that I don’t want much better multi-core CPU scalability, I do, as more shards are still harder to manage than less shards, it just isn’t as important as better sharding functionality.  I absolutely need the ability to shard; I don't absolutely need the ability to effectively use every CPU on a server with one shard. 

On another note, I’ll be giving a presentation on table partitioning at the MySQL conference.  This won’t be a basic presentation on how to implement partitioning, instead I’ll be highlighting scenarios where partitioning will degrade performance and cases where it improves performance.  I’ll show examples where, in spite of degraded performance for some tasks, it still might make sense to implement it.  An analogy here would be indexes, where adding an index can speed up some selects at the expense of slowing down inserts and some updates.

Infobright Review – Part 2

First, a retraction, it turns out that the performance problem with datatimes in the previous article wasn’t due to high cardinality (I speculated too much here), but due to a type conversion issue.  From a helpful comment from Victoria Eastwood of Infobright (a good sign for a startup), the Infobright engine considered ‘2001-01-01’ to be a date, not a datetime, and it couldn’t do a conversion to a datetime.  Instead it pushed the date filtering logic from the Infobright engine to MySQL.  Effectively, the slow queries were a table scan.   The solution is to add the 00:00:00 to the dates to make them datetimes.  

With that in mind, here are some much better numbers for Infobright.   For Infobright this query took 0.05 seconds. 

1) Select sum(unit) from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-05-01 00:00:00'

This compares very favorably to 4 seconds for InnoDB if the data is cached and 8.8 if the data needs to be read from disk. 

In doing further tests I found some interesting patterns one might not expect coming from a row database background.  This more selective #2 query took 0.27 seconds for Infobright, longer than the less selective query #1.  Still 0.27 is fast enough.  The odd part is a row based database is generally faster, or at least the same performance, the more selective the query.  For Infobright the exact opposite occurs, at least in these cases I explored.   For this more selective query, InnoDB took 0.047 seconds with a cold cache (all following queries will be using a cold cache unless otherwise noted) compared to that 0.27 for Infobright.  It seems that the that Infobright is faster than InnoDB for less selective queries, and InnoDB is faster for more selective queries. 

2) select sum(unit) from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-05-01 00:00:00' and productId = 199824

This more selective query #3 took longer still, at 0.99 seconds for Infobright, but only took 0.047 seconds for InnoDB, the same time as query #2.  

3) select sum(unit) from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-05-01 00:00:00' and productId = 199824 and customerId = 418845

The same pattern becomes more evident with double the time range for the queries, with two months of data compared to just one.   Query #4 took 0.07 seconds for Infobright whereas InnoDB took 19.9 seconds, far longer.  

4) select sum(unit)  from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-06-01 00:00:00'

At 1.25 seconds, adding another criteria noticeably slowed down the performance from query #4 for Infobright.  At 0.061 seconds, InnoDB was much faster than the less selective query #4,

5) select sum(unit) from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-06-01 00:00:00' and productId = 199824

At 1.99 seconds, this more selective query #6 was again slower than #5 for Infobright.  Following the trend, InnoDB was even faster at 0.048 seconds.  

6) select sum(unit) from Sale where purchaseDate >= '2001-04-01 00:00:00' and purchaseDate < '2001-06-01 00:00:00' and productId = 199824 and customerId = 418845

But there is another difference between InnoDB and Infobright.  The below query #7 takes 0.06 seconds for InnoDB, and takes 11.5 seconds for Infobright with a cold cache (I restarted the database in both cases).  Once the cache was warmed up Infobright took 2.2 seconds to run and retrieve one row.  As there are 120 million different orderIds in the Sale table, perhaps my speculation that Infobright does worse the higher the cardinality of a column wasn't that off.  

7) select * from Sale where orderId = 1

At least for the cases I examined, InnoDB is better for more selective queries, at times far better, and Infobright is better for less selective ones, at times far better.  This assumes defining good indexes for InnoDB, work which Infobright does not require.  Which one to use for reporting?  It really does depend on the nature of the workload.   

Having said that, Infobright behaves quite differently from InnoDB and I know there is much I'm missing as I only put in a few hours of work.   I would try your own workload and come to your own conclusions.