SQL Server: High performance inserts

http://www.altdevblogaday.com/2012/05/16/sql-server-high-performance-inserts/

by dmoney67 • 14 years ago

42 32 14 years ago

One optional 5th revision that can get even better performance for particularly large inserts: SqlBulkCopy[1]

In my experience, it tends to pull ahead of table-valued parameters somewhere between 5 and 10,000 rows. On the other hand, the way it does locking makes it less suitable for concurrent scenarios.

[1]http://msdn.microsoft.com/en-us/library/system.data.sqlclien...

balakk• 14 years ago

Another issue is Bulkcopy sucks for error reporting and txn management. TVPs are an acceptable compromise for most situations; if only they would let us create a TABLE type based on an existing schema.

emn133• 14 years ago

How are they any worse for error reporting?

Also, unless I'm mistaken, if you want to, you can enroll a bulk copy into a custom transaction for maximum control and essentially replicate the table-valued parameter behavior.

emn133• 14 years ago

The locking options can be controlled - http://msdn.microsoft.com/en-us/library/system.data.sqlclien... Is there some other locking issue?

bunderbunder2• 14 years ago

None of the locking options work for concurrent bulk inserts if there's an index on the table. Trying to do so invites a mess of deadlocking.

lusr• 14 years ago

And on SQL Server < 2008 where TVPs aren't an option and the data set is large enough (I was importing millions of rows from user uploaded CSV files) I've had good experiences building chunks of XML (couple thousand rows at a time) on the client side and shipping that off to the client as a single XML parameter.

Although, in practice, when importing complex data in real-time there's a lot of experimentation required to determine where the real bottlenecks are and this article skims the surface of the complexity involved in a real import.

For example, do you do your lookups inside your stored procedure or does your application layer deal with all of that and send off pre-prepared data? If the latter, you reduce the amount of time the table is locked whilst adding rows since there are no joins and lookups in the insert, and this may be more performant overall than the alternative.

You can also reduce the duration of any transaction lock you may need to hold for importing all the data which ultimately makes the system more responsive during large imports. Furthermore the "prepare-first" approach makes error reporting more practical since it's generally easier to do and report validation in the application layer.

j_baker• 14 years ago

    -- No longer has a DEFAULT constraint; this means we have to insert the date
    -- ourselves, but it reduces the work on SQL server
    kill_time datetime

My gut reaction is that this is unnecessary micro-optimization. Does making SQL server come up with the date and time really make that big a performance difference?

emn133• 14 years ago

Not just that; it might well be slower. It's definitely possible to saturate a network connection with a bulk operation like this; less wire overhead matters.

I suspect it's mostly just necessary to permit the delay between data generation and storage.

Cushman• 14 years ago

I think you misread the code-- the datetime is still being generated in SQL, in the stored procedure, as GETDATE(). Nothing extra is going down the wire. The point is to remove the default constraint, and thus the check for null on every insert.

May still be unnecessary, but seems like a fair optimization.

pikewood• 14 years ago

It's interesting to see how the code pulls it off. If you run Profiler while this runs, you'll see a table variable being created, and then an insert statement for each row into that temp table. My guess is that the main performance gain comes from inserting into a table that doesn't require checks against constraints, etc., but eventually the overhead for all those individual inserts means that SqlBulkCopy wins out.

And, the need to define your UDTs is a bit of a pain. But to prevent a glut of UDTs in the db, I've tried to make common UDTs that would commonly be used, such as

  CREATE TYPE [dbo].[udt_IntTable] AS TABLE(
	[Item_Key] [int] NULL
  )

One more point: it's probably best practice to use the TypeName property when using these, as it's required when using a UDT against a parameterized statement that's not a sproc.

  param.TypeName = "item_drop_bulk_table_rev4";

gnaritas• 14 years ago

> My guess is that the main performance gain comes from inserting into a table that doesn't require checks against constraints

My guess would it's because inserting into table variables isn't logged; doubt it has anything to do with constraints. This turns all those individual statements into a single bulk insert, a single larger log operation instead of many small ones.

orcadk3• 14 years ago

There are quite a lot of inaccuracies in this post, as well as hints that don't really mean much.

First up, the non-SQL-injection code doesn't specify the parameter types. As such, each length of monster_class will result in a different execution plan being compiled and cached by SQL Server. Granted, there's only 50 different kinds - but just specify the parameter types to avoid it.

Sure, you could setup a VLAN, but we're lightyears from having that affect our performance at this point.

Parser overhead is not an issue as long as parameter types are specified. The plan will be compiled and cached. The only advantage to stored procedures at this point is reduced network data as well as management.

Page faults - Sure, but for inserts the data will most likely be in memory already as we're writing new data. And if there's enough data, the relevant b-tree pages will most certainly be hot.

Varchars - Granted, avoid them unless you're storing string data. I have no idea where the "hundreds of bytes of of memory" comes from. SQL Server caches all data as pages, exact copies of what's stored on disk. Variable-length varchars have an overhead of exactly two bytes + 2/#varlengthCOlumns in total.

Disk IO - The transaction log is never merged with the main data file. SQL Server uses write ahead logging, writing to the log first while modifying the data pages in memory. Now and then a checkpoint is run, writing all the dirty pages to the main data file. A log is only "merged" with the data file if the server has crashed and recovery is being run automatically during startup.

"This increases the number of rows you can insert before SQL Server succumbs to memory pressure" - Since the article only deals with inserts, why are we talking about memory pressure? You'll get disk contention for sure, but memory won't realistically be an issue unless you're reading as well.

There's an alternative to keeping the potentially long running transaction open - avoiding both SqlBulkCopy and transferring datatables. The BCL has a CommandSet notion that utilitizes a TDS feature that allows batches to be streamed to SQL Server. Unfortunately you'll have to use reflection to get at it (https://github.com/ayende/rhino-etl/blob/master/Rhino.Etl.Co...), but the code has been stable for several releases and is very simple to interface. It basically allows you to create normal SqlCommands and just send a large batch of them at once, all being executed atomically with minimal locking.

Finally, with the micro optimizations being mentioned, I'm surprised he doesn't recommend to "SET NOCOUNT OFF".

Overall, some good points with some inaccuracies. I have a hard time with the closing statement "As you can see, you don’t need to abandon SQL Server to get massive performance improvements.". These are decent improvements, but this is such a simple scenario that I find it pretty optimistic to let this define whether to abandon SQL Server or not. The RDBMS problem is rarely with inserting enough data, but with reading data at the same time while avoiding locking.

leif4• 14 years ago

"and if there's enough data, the relevant b-tree pages will most certainly be hot"

That is not the case if you have indexes. Most secondary indexes are high-entropy (if they weren't, you probably wouldn't be storing them, is an emotional proof), so inserts on larger-than-memory data sets almost always incur a disk I/O, in a b-tree, at least.

Shameless plug: I have a talk about how to deal with this: http://www.youtube.com/watch?v=q6BnG74FZMQ

DenisM2• 14 years ago

In fact inserts almost never incur a disk IO, not until the transaction commits. Even then the IO is sequential, in the log file, not random in the mdf file.

The mdf pages will get marked "dirty" at the time of insert, and they will get flushed to disk eventually - either when checkpint time comes, or if memory gets scarce and the dirty pages get evicted.

It might seem that postponing the write until checkpoint/eviction is not meaningful, but it actually serves a purpose - for one, the transaction is able to finish and return confirmation to the user very quickly, and then if the same page is touched twice (or more) between two flushes, all but the first touches are "free".

orcadk3• 14 years ago

I don't agree that most secondary indexes are high cardinality. Some definitely are, but there are plenty of indexes with low cardinality, simply used for quicker scanning of those relevant values.

First up, in his example we're talking about 10k rows with a size of 35 bytes. At 8050 bytes of available space per page, that gives a fan-out of 8050/35 = 230. 10k/230*8kb = 350kb. At 350kb of inserts (hobt data, he doesn't go into secondary indexes), memory is completely irrelevant - the only force going on here is latency on writing the log to disk.

If we had a huge data set (as in, did not fit in memory, at all) with high cardinality - sure, we'd have a lot of cold leaf level pages. With no further info on his case, I can only assume most of the hobt and secondary indexes will fit in memory. At worst we'll have to read a cold leaf level page into memory to perform the addition in-memory.

As is there's no mention of even a clustered index, causing all of this to be heap inserts which is arguably one of the fastest insert methods there are (barring certain very special cases).

simplezeal• 14 years ago

I agree with most of your points.

I have to give it to the author though for a genuine attempt at one of the important things when dealing with databases. There is lot of room for performance improvement in application code, before you start blaming the platform for not scaling.

chris_wot2• 14 years ago

Another thing is that he doesn't need to use a stored procedure, he could use prepared sql statements.

In regards to your last sentence, he could change the isolation mode if he didn't care so much about consistency.

orcadk3• 14 years ago

He already is using prepared statements, that happens under the hood when you use parameterized queries. Problem is the lack of type definition (in the first revision), causing a new compilation for each length of the varchar column.

koide2• 14 years ago

For really massive data inserts, bcp is the way to go. In general avoid bulk insertions in code, if at all possible. If that's not an option then go for SqlBulkCopy or other similar approach.

sams993• 14 years ago

bcp and SqlBulkCopy are using the same underlying db mechanism.

koide2• 14 years ago

I read somewhere I cannot find right now that bcp is faster than SqlBulkCopy. My experience tends to match this, but that very well might be due to other factors (network, wrong coding, different default values between SqlBulkCopy and bcp.)

Do you have a source to back your statement up? That way I could stop looking for mine :)

DonnyV• 14 years ago

Looking at this reminds me why I'm happy to work with MongoDB. I do not miss TSQL.

sirfried• 14 years ago

Well for the money I'd pay in licensing this, I can just pay for more servers and use MySQL or PostgreSQL. Not that those two are slow at all.

pwaring• 14 years ago

You could, but that solution isn't of much help to someone with an existing .NET/SQL Server application. Making a few tweaks is probably a lot cheaper than a wholesale change to a different database server.

Umox• 14 years ago

Uhm, or just use a document database without all that ugly procedural code that ruins testability and obstructs OO patterns and practices.

SQL is quickly becoming a legacy product.

SQL Server: High performance inserts

Related Stories