Skip to content

Delayed Durability in the wild…

Background

We have recently been working on large data migration project for one of our clients and thought I would share how Delayed Durability helped us overcome a performance issue when the solution was moved to the client’s Development domain.

I won’t go into details of the project or the finer detail of our proposed solution as I have plans to put some more content together for that but in short the migration of the data was to be run by a (large) number of BIML generated SSIS (Child) packages for each table to be migrated, derived from a meta data driven framework with each stage being run by a master package, all of which run by a MasterOfMaster packages.

To maximize throughput, utilise as much processing power as possible, reduce the time it would take to run the migration and control the flow we built a series of sequence containers each running it’s own collection of Child Packages. We built the framework in such a way that these could be run in parallel or linear and each master package could contain as many containers (no pun intended) of child packages as required. This also allowed us to handle the order that packages were run in, especially those with dependencies whilst keeping the potential for parallelising (is that a word? No idea but I like it) the whole process as much as possible. Leaving the MaxConcurrentExecutables property to -1 mean’t we could push the processing to run up to 10 packages at ones due to the VM having 8 cores (on Integration, 4 cores on Development) and this value of -1 allows the maximum number of concurrently running executables to equal the number of processors plus two.

An small example of how the MasterOfMaster and a Master Package for a stage looked is shown below:

Each container number could have Parallel and/or Linear Containers and both must succeed before the next Container level can start.

NOTE that this is just an example representation, naming conventions shown do not reflect the actual solution.

Problem

During development and initial testing on our own hardware, we had the migration at the time running at ~25minutes for around 600 packages (ie. tables) covering (what we termed) RawSource–>Source–>Staging which was well within the performance requirements for the stage that development was at and for what was initially set out. The rest of this blog post will hone in specifically on Source–>Staging only.

However, once we transferred the solution to the clients development environment things took a turn for the worse. In our environment we were running VMs with 8 cores, 16GB RAM and utlising SSDs. The client environment was running SQL Server 2016 Enterprise on VMWare vSphere 5.5, 8 vCPUs, 32GB RAM (for Integration, Development was half this) but the infrastructure team have done everything in their power to force all VMs onto the lower tier (ie. slow disks) of their 3-PAR SAN and throttle them in every way possible, just to make things more of a challenge. Even though the VM’s themselves were throttled we were confident that we wouldn’t see too much of a performance impact, especially as this was only a subset of the processing to be done so we needed it to be quick and it will only ever get longer and longer.

How wrong we were. On the first run the processing (for Source–>Staging) took around 141 minutes, yes you read that right, a full 116 minutes longer than the whole process took on our hardware! Wowza, didn’t see that one coming. I won’t delve too much into the investigations as again that will be saved for another blog post but essentially we were seeing a huge amount of the WRITELOG wait type since moving to the client environment. We believed the reason for this was due to the significant amount of parallel processing (running of SSIS packages in parallel loading to the same DB) we were doing and the SAN didn’t seem to be able to handle it. One other thing to note, due to truncations not being flagged as error’s in OLE DB Destination fast load data mode access, some of the packages that weren’t a direct copy where we knew the schema was exactly the same were run in non-fast load, ie row-by-row which puts additional stress on the system as a whole.

I will be blogging at a later date regarding how we managed to get everything running in fast load and handle the truncation via automated testing instead.

Solution

Enter Delayed Durability.

I won’t enter into too much detail regarding what this is or how it specifically works as this has been blogged by many others (Paul Randal, Aaron Bertrand to name just a couple) but my favourite description of delayed durability is comes from the msdn blogs and they refer to it as a “lazy commit“. Before you ask, yes we understood the issues of implementing such a change but the migration process was always a full drop and reload of the data so we didn’t care if we lost anything as we could simply run the process again.

Setting delayed durability at the database level we were able to control which Databases involved in the process we wished to have this without altering the BIML framework or code itself to handle it at the transaction level. By simply applying this to the Source and Staging databases we reduced the processing time from 141 minutes to 59 minutes. This wasn’t exactly perfect but shaving more than half the time off with one simple change and pushing the WRITELOG wait stat way way down the list was a great start.

As a side not, we have managed to get the processing from ~59mins to ~30mins without changing the VM/hardware configuration but I will leave that for another post.

Proof

When I first set out with this blog post it was only going to be a few paragraphs giving an insight into what we did however, I thought that all this would be pointless without some visualisation of the processing both before and after.

Row-by-Row with no Delayed Durability

We needed to get a baseline and where better to start than capturing the metrics through SentryOne and using Adam Mechanic’s spWhoIsActive we can see what I was talking about with the WRITELOG wait stat:

spwhoisactive_output

Granted the wait time themselves was relatively low, these were apparent almost every time we hit F5 and running our wait stat scripts in was in the top 3. A sample of the processing indicating this wait stat can also be seen below:

waitstats_writelog

As stated previously, overall the Source–>Staging process took 141 minutes and the overall processing from SentryOne PA was captured:

so_fullprocess_output

Row-by-Row with Delayed Durability

So when we ran the same process with Delayed Durability we can see straight away that the transactions/sec ramp up from ~7000 to ~12500. Top left shows without Delayed Durability, bottom left with Delayed Durability and right shows them side by side:

The overall process for Source–>Staging took only 59 minutes. I’ve tried to capture the before/after in the image below, the highlighted section being the process running with Delayed Durability forced:

SO_FullProcess_withDD_Output.png

You can see from this the drastic increase in Transactions/sec and reduction in Log Flushes.

Two package execution time examples (trust me that they are the same package) showing that with Delayed Durability the processing time was only 43% (166sec down to 72sec and 991sec to 424sec) ) of that without Delayed Durability set. Apologies for the poor image quality….

To me that is a huge reduction for such a simple change!

Conclusion

So should you go out and apply this to all your production databases right this second? No, of course you shouldn’t. We applied this change for to fix a very specific problem in an isolated environment and were willing to take the hit on losing data if the server crashed – are you, or more importantly your company willing to lose that data? I’m taking an educated guess that this will be a no but for certain situations and environments this configuration could prove to be very useful.

Links

Why would you never use SSIS Fast Load…?

We all know that if you want SQL Server to push data into a table then you want to batch the inserts / use a bulk insert mechanism but is there a time when performance isn’t everything?

Background

Although it has its critics, SSIS is a very powerful tool for Extracting, Transforming and ultimately Loading data from and to various systems. I kind of have a love / hate relationship with SSIS, I love it but it seemingly hates me with a passion.

During a recent data migration project we had a series of packages using a stored procedure as the source and a SQL Server table as the destination. By using the OLE DB Destination task you have a series of options Data Access Modes which can provide various additional configurations. I won’t delve into all of these but have a look at the msdn link provided at the end for further information.

The ones I want to concentrate on are:

  • Table or view
  • Table or view – Fast Load

In short, fast load does exactly what it says on the tin, it loads data fast! This is because it is optimised for bulk inserts which we all know SQL Server thrives on, it isn’t too keen on this row-by-row lark.

Problem

Now, I won’t be providing performance figures showing the difference between running a package in fast load compared to row-by-row, this has been done to death and it is pretty much a given (in most cases) that fast load will out perform row-by-row.

What I do want to bring to your attention is the differences between the two when it comes to redirecting error rows, specifically rows that are truncated. One of the beauties of SSIS is the ability to output rows that fail to import through the error pipeline and push them into an error table for example. With fast load there is a downside to this, the whole batch will be output even if there is only 1 row that fails, there are ways to handle this and a tried and tested method is to push those rows into another OLE DB Destination where you can run them either in smaller batches and keep getting smaller or simply push that batch to run in row-by-row to eventually output the 1 error you want. Take a look at Marco Schreuder’s blog for how this can be done.

One of the issues we have exerienced in the past is that any truncation of a column’s data in fast load will not force the package to fail. What? So a package can succeed when in fact the data itself could potentially not be complete!?! Yes this is certainly the case, lets take a quick look with an example.

Truncation with Fast Load

Setup

I have provided a script to setup a table where we can test this. I will attempt through SSIS to insert data which is both below and above 5 characters in length and show the output.

USE tempdb;
GO

DROP TABLE IF EXISTS dbo.TruncationTest;
DROP TABLE IF EXISTS dbo.TruncationTest_error;

CREATE TABLE dbo.TruncationTest
(
TruncationTestID INT IDENTITY(1,1),
TruncationTestDescription VARCHAR(5)
)
GO

CREATE TABLE dbo.TruncationTest_error
(
TruncationTestID INT,
TruncationTestDescription VARCHAR(1000) --Make sure we capture the full value
)
GO

This code will set up 2 tables, one for us to import into (TruncationTest) and another to capture any error rows that we will output (TruncationTest_error).

I set up a very quick and dirty SSIS package to run a simple select statement to output 3 rows and use the fast load data access mode:

SELECT  ('123') AS TruncationTestDescription UNION ALL
SELECT  ('12345') UNION ALL
SELECT  ('123456789');

The OLE DB Source Editor looks like this:

sourceeditor

the OLE DB Destination data access mode:

desteditor_fastload

Finally, this is how the package looks:

package_fastload

Note the truncation warning. This is easy to see when viewing a package in Visual Studio, not so easy to pick up when you are dynamically generating packages using BIML.

Let’s run it……

package_fastload_success

Great, 3 rows populated into the TruncationTest table, everything worked fine! So let’s check the data:

SELECT * FROM dbo.TruncationTest

results_1

Eh? What happened there???? Where’s my ‘6789’ gone from row 3???

From this example you can see that the package succeeds without error and it looks as though all rows have migrated entirely but by querying the data after the package has completed you can see that the description column has indeed been truncated.

Let’s try the same test but changing the Data Access Mode to non-fast load (ie. Row-By-Row)

Truncation with row-by-row

In this example you can see that the row with truncation is in fact pushed out to the error pipeline as you would hope and expect.

desteditor_nonfastload

package_nonfastload_success

We now have 3 rows being processed but one row pushing out to the error pipeline which is what we would expect and hope for.

Let’s take a look at the output:

SELECT * FROM dbo.TruncationTest ORDER BY TruncationTestID
SELECT TruncationTestDescription FROM TruncationTest_error

results_2

The results highlighted in red are those from the fast load, in green are the results from the row-by-row indicating that the error row was piped out to the error table.

Solution(?)

You have a few different options here:

  1. Not really care and push the data through in fast load and suffer the concequences
  2. Run in row-by-row and suffer the performance hit
  3. Amend the OLE DB Source Output to be the same length as the destination column and redirect error rows from there.
  4. Probably loads of others involving conditional splits, derived columns and/or script tasks
  5. Apply option #1 and make sure that relevant (automated or otherwise) testing is applied

During the recent data migration project we were involved in we chose option #5. The reasons for this are:

  1. We wanted to keep the BIML framework, the code and the relevant mappings as simplistic as possible
  2. Performance was vital….
  3. …..but more importantly was the validity of the data we were migrating

We already had a series of automated tests setup for each package we were running and table we were migrating and we had to add to this a series of additional automated tests to check that no data itself was being truncated.

NOTE: Option #4 was also a very valid choice for us but due to the nature of the mapping between source and destination this was not something that was easily viable to implement.

I will leave the how we implemented these test this for another blog post🙂

Conclusion

Taking a look at the error redirect in the OLE DB Destination we can clearly see that Truncation is greyed out and no option is provided so I have to assume that it simply isn’t an option to configure it here.

errorredirect

I used to have a link to an article which mentions that truncation cannot be deemed an error in a bulk import operation via SSIS due to the mechanics of how it all works but for the life of me I cannot find it😦. I am hoping someone who reads this will be able to provide me with this but for now I will have to draw my own conclusions from this. The closest thing I can find is an answer from Koen Verbeeck (b|t) in an msdn forum question where he states:

The only thing you get is a warning when designing the package.

You get truncation errors when you try to put data longer than the column width in the data flow buffer, i.e. at the source or at transformations, but not at the destination apparently.

What I still don’t understand is why in tSQL you will get an error when trying to “bulk insert” (loose sense of the term……ie. using an INSERT….SELECT) data that will truncate data but SSIS does not. Hopefully someone far cleverer than me will be able to shed some light on this!

The idea behind this blog post was not to focus too much on the importance of testing any data that is moved from one place to another but I wanted to highlight how easy it is to believe that what you are migrating is all fine n dandy because the SSIS package told you so but in actual fact you could be losing some very very important data!!

You have been warned😉

Links

Windows 10 Upgrade on Surface Pro 3 – Sending email issue FIXED

windows-10

I very rarely blogabout anything other than SQL Server but felt that with a lot of the SQL Community using SP3’s that this may actually be helpful to some.

After playing around in a VM for a while now I decided to upgrade my SP3 to Windows 10 which comes as a free upgrade. I was surprised that after making the schoolboy error of believing my charger was actually charging the unit during the upgrade and the battery running flat 30% through the upgrade, it fired back up and started off from that point with no issues whatsoever! I honestly thought that was it and it would be a factory reset.

I won’t go on about what’s changed, what I like and what I don’t like, you can read plenty of that around on the internet and it is out of the scope of this blog post.

However, I did discover one issue which I know several others have had when upgrading their SP3. After the upgrade I could no longer send emails via (IMAP) outlook. I could send via every other device so knew it was specific to the upgrade. I managed to find a couple of others on the windows forums with the same issue and they provided the solution which I thought I would share.

The fix is actually very simple, all you need to do is open a command prompt, run sfc /scannow and wait ~10mins.

The System File Checker will scan your Windows system files for corruption and attempt to repair them, below is the output from my run:

sfc

I won’t go into details of the log file but the repaired file was related to the “Multilingual User Interface” (mui) files, which are translation files used to support different languages within windows. Mine in particular was mlang.dll.mui.

So, after a mild panic the fix was relatively straight forward……once I knew what the problem actually was!

SQLNorthEast Usergroup 2015 dates announced (preliminary)

Mike and I have been extremely busy over the Xmas period and we’ve finally sorted dates for our 2015 instalment of the SQLNorthEast SQL Server UserGroup (@SQLNE) in Newcastle. The great news is that after much negotiation we have managed to get agreement in principal to use the same venue for our events which is fantastic news!

Please see www.sqlne.com for info on our next meeting and registration. Due to sqlpass website restrictions we cannot display all the dates for 2015 but a quick search on Eventbrite will give you the relevant details.

The dates are as follows:

Feb: Tue 10th (with Chris Adkin – double session)
March: Tue 24th (with Erin Stellato from SQLSkills and Peter Shaw)
April: Tue 28th (with Neil Hambly)
June: Tue 2nd (with Steve Powell)
July: Tue 7th (with Annette Allen)
Sept: Tue 8th (TBC)
Nov: Tue 24th (TBC)

As you can see we have already lined up a number of fantastic speakers including a special remote session from the world class sqlskills Principal Consultant Erin Stellato!

2014 was a great year for us and due to the success of our second SQLRelay we’re going to be sending out a survey to find a bit more information about the needs and wants of our delegates to help us set our content for 2015. We have a very mixed bag of experience as well as SQL Server areas so any help or ideas we can get is a great help. If you wish to complete this now then you can find it at SurveyMonkey.

If you have any thoughts or ideas regarding how we can improve then please get in touch via email or twitter and we’ll endeavour to incorporate it.

Merge csv files – quick PowerShell snippet

During a bit of work I’ve been doing this evening for SQLRelay 2014 I used something I have in my arsenal of PowerShell scripts which I thought I’d share because I love it’s simplicity🙂 It’s nothing big and fancy but something that is extremely useful. Tasked with merging a large number of csv files there (as always) is a quick and easy way to do this with PowerShell:

Get-ChildItem *.csv |
ForEach-Object {Import-Csv $_} |
Export-Csv -NoTypeInformation WhateverYouWantToCallTheFile.csv

There are ways to make this a little more dynamic which I will update the post with in the coming weeks….

2013 in review

The WordPress.com stats helper monkeys prepared a 2013 annual report for this blog.

Here’s an excerpt:

The concert hall at the Sydney Opera House holds 2,700 people. This blog was viewed about 12,000 times in 2013. If it were a concert at Sydney Opera House, it would take about 4 sold-out performances for that many people to see it.

Click here to see the complete report.

Output SQL Server data from multiple tables to Tab Delimited text files using Powershell

I had a request this morning for something I though was actually very simple:

Client: “Can you extract all data for these particular tables including column headers to a tab delimited .txt file?”
Chris: “Sure, no problem, I’ll just run bcp querying sys.tables using a COALESCE loop to output the statements”
Client: “Top stuff, let me know when it’s done”

So, away I went generating my script which took a matter of minutes and run it…….where’s the column headers? Bugger, forgot that bcp doesn’t output column headers without doing some funky stuff by creating a header record in a separate file and merging that with the file of data.

With this in mind I knew creating a SSIS package (or using export data to generate – very manual unless I delved into the realms of BIML) could do this but I thought I’d have a look at powershell invoking sqlcmd.

Again, this all seemed to be going very well until I came to outputting the data to a tab delimited .txt file. As far as I’m aware Powershell does not have an Export-Txt so I had to look into how I can use the Export-Csv to actualy output to .txt tab delimited as opposed to comma separated and found the parameter -delimiter “`t” – Excellent!!! Added this in and run the script…………and the first row consisted of “#TYPE System.Data.DataRow” – wft!?!?!?!?!

Quick search on my search engine of choice showed that there is a parameter that you can pass in to remove this from the export -NoTypeInformation.

Run it again with -NoTypeInformation and everything worked as expected apart from all column headers and data had quotes (“) around them which was not part of the requirement. Unfortunately (as far as I know) there is no switch, parameter or the likes that does this so I had to change the Export-Csv to ConvertTo-Csv and run a Replace on ‘”‘ with ” which managed to do the trick.

I’ve included the script below which can be tailored to your needs:

$server = 'ServerInstanceHere'
$database = 'DBNameHere'
$path = 'c:\work\ToDelete\'
$query = "SELECT name FROM sys.tables WHERE name in (
 'TableNameHere_1',
 'TableNameHere_2'
 --etc etc
 )"
$queryToOut = "SELECT * FROM $TableName"

#Get list of table names to output data
$Tables = invoke-sqlcmd -query $query -database $database -serverinstance $server
foreach ($Table in $Tables)
{
 $TableName = $Table["name"]
 write-host -ForegroundColor Green "Creating File $TableName.txt"
 invoke-sqlcmd -query $queryToOut -database $database -serverinstance $server | `
 #Convert as opposed to Export to replace quotes if required
 ConvertTo-Csv -NoTypeInformation -delimiter "`t" | `
 ForEach-Object {$_ -Replace('"','')} | `
 Out-file $path$TableName.txt
 #Export-Csv -NoTypeInformation -delimiter "`t" -path $path$TableName.txt
}

Apologies for the formatting but the powershell script tag doesn’t seem to format it the way I’m wanting it to😦