Monday, February 17, 2020

Another note on versions used in this blog

Hello readers! Last year I made a post about the SQL Server versions I use for this blog, and it's time to post an update.

I'm a Mac user, so I've historically had to run SQL Server in a Windows virtual machine. Thanks to Microsoft's expansion into the Linux world, this has opened up my options quite a bit! As time has passed, I've largely stopped using SQL Server on Windows for my lab environment, focusing instead on the Linux version. I'm currently using SQL Server 2019 CU2 on Ubuntu 18.04.4 Server LTS running in a Parallels Desktop hypervisor (VirtualBox is a great free alternative, and runs on a wide variety of platforms). I also use Docker containers for ephemeral instances. I use Azure Data Studio and Visual Studio Code to do my management and dev stuff.

As such, please assume unless stated otherwise that everything in this blog is based on SQL Server 2019 for Linux, on whatever CU is current at the time of publication (though it might be the previous CU if a new one has been released just before I publish). Screenshots are going to generally be from Azure Data Studio. Once in a while I'll do something in SSMS or on a Windows-based instance, but I to call those out specifically.

Please note that SQL Server is officially supported on Ubuntu 16.04, though. I tried to get it running on 18 as a lark (it started life as a SQL Server 2017 instance, upgraded in place to 2019, and while it's doable, there are some hoops to jump through), but I don't do any production work here.

You've got lots of options these days for running SQL Server on non-Windows platforms! Soon I'll be learning Azure SQL Database, so that'll be another new frontier!

Sunday, February 16, 2020

Scaling is hard

Hello readers! I apologize for the lack of posts for the past... (checks calendar) six months... *cough* I will resume my series of posts on change tracking in the near future! In the meantime...

Enterprise scaling is hard. Anyone who says differently is selling something. Due to the incredible public visibility (to say nothing of the... investment... some of the fans bring), scaling in video games adds a miles-thick layer of difficulty. Recently, Bungie has had a few significant bugs crop up in their flagship title, Destiny 2. These bugs affected player inventories, and led to unplanned downtime while fixes were implemented. Twice. (Insert your choice of 'oof' meme here.) Bungie has published a good post-mortem that, while understandably light on the technical details, gives some insight into what they were doing and how things broke.

This incident called to mind something I worked through. It's absolutely not the same level of issue, but it's illustrative of the interesting things you see at scale. A previous employer's business requirements meant we had servers close to just about every major city on the planet, and they were kept fairly busy. Due to the nature of the business even a server that wasn't busy had to be very responsive, and a busy server that got slow was a nightmare for customers.

One server began periodically getting slow. REALLY slow. A proc that averaged a worker time of under a millisecond (props to the developers there: the tables were well-normalized, the queries well-written, and they understood indexing so the indexes were sensible as well) started taking over a second to respond. Our initial thought was that it had generated a bad plan by dumb luck (due to the nature of the data and how it changed over time, plans got regenerated a few times a week following stats maintenance, so far so normal), so we started by comparing the plan on this server to the plans from a sampling of its compatriots. The plan was indeed different (and bad!) on this one, so we gave the business a heads-up that we were going recompile that proc, we'd probably see a long execution while the new plan was generated, and so on. Recompiling worked a treat, the new plan was good, and matched the good plans on the other servers, so we were happy, the business was happy, everybody was happy.

A few weeks later, this one server started acting up again. Meet the new bad plan, same as the old bad plan. But only on this one server. We looked at a different sample of other servers of the same type - all consistent among themselves, and they all had generated the same good plan. So, again with the recompile, which made everything happy again. Once again the new plan was good, matching the good plan everywhere else.

We didn't wait for the third time to assume that it wasn't just dumb luck causing the bad plans. Parameter sniffing is fun, after all. One of the fun parts here was that even though the plan WAS using the desired index, it was just... wrong. It's worth noting that all the servers of this type had the same source tables replicated down to them. No local writes, just reads, and the data sets were identical.

So why was this lone box coming up with this suboptimal plan even using the correct index? Well, one common thread was that the bad plans were generated from one parameter value. The application these servers supported did a nightly test, and each server had a short list of test values that they'd cycle through. The tests were server-specific so no two servers would use the same test records. The server tests were done during slow hours for that region, and just so happened to happen after index maintenance was done. Every so often, the old plan would be cleared after maintenance and we'd get the bad plan when the right test parameter was passed in.

I started digging into the index. DBCC SHOW_STATISTICS is tremendously useful for understanding what a given index looks like under the hood. The histogram in particular was helpful, as it allowed me to tell where in the index the bad parameter was found. I pulled copies of the various objects related to the issue down into my sandbox and set to work on it.

After ensuring I could reproduce the problem at will, I set to checking the parameters used to generate the good plans. Some of the values corresponded to test records, some corresponded to customer data, but interestingly they all were found in different steps in the index. I noticed that the steps weren't all of identical size - the step with the bad parameter had a larger range than the other steps. Out of curiosity, I tried passing some other parameters from the same step as the parameter that produced the bad plan, and those parameters also caused the same bad plan to be generated. Similarly, other values from the steps that contained test values that generated good plans would generate good plans.

So far, so good. I did some inquiring and found that the bad parameter value was the only test value in that index step, but there was always the chance that a plan could be generated using some other parameter from that step if the plan fell out of the cache for whatever reason.

My next step was to see if changing our indexing strategy would help. The table had grown a lot over time since the application was rolled out, and maybe we'd hit the point when rethinking things was the right thing to do. A discussion of the nuances of index design is far outside the scope of this post, but as with most things the online documentation covers it pretty well, and of course many smart people have spilled a lot of virtual ink on the topic; your search engine of choice is your friend here.
I tried a number of things both on my own and in consultation with folks who know a lot more about indexes than I do, but nothing helped. While I was able to get the optimizer to consistently generate a plan that was slightly better than the bad plan regardless of the parameter passed, it was still substantially worse than the good plan. The solution wasn't to be found in the indexes, to be sure.

In the end, we wound up creating a job that checked the stored procedure's performance and would force a recompile if the maximum worker time was over a certain duration. We scheduled the job to run every few minutes, and this took care of the problem sufficiently well. We shortly deployed this to all the relevant SQL instances, and it turns out that it periodically helped prevent the problem on other servers as well over time. I had an opportunity to talk with some folks from Microsoft later, and they agreed that under the circumstances the recompile job was the best solution.

Scale gets strange at times. Things that you never encountered before start happening more and more frequently, and sometimes you have to approach the problem differently than you think.

Saturday, August 17, 2019

Tracking Data Changes - Change Tracking, Part 2

In my last post, I showed you how to enable, configure, and disable change tracking at the database level. Very good, as far as that goes, but you're not going to get very far without tracking some actual data changes!

To recap, enabling change tracking at the database level is very simple:
ALTER DATABASE <database name here> 
SET CHANGE_TRACKING = <ON | OFF>
(CHANGE_RETENTION = <positive integer> <DAYS | HOURS | MINUTES>, AUTO_CLEANUP = <ON | OFF>);

(Note on the code snippets: I'm using Markdown for this post to simplify my blogging workflow; please let me know if you'd rather I keep using Gists for code samples)

If you don't provide a retention period, SQL Server's default is 2 days. Auto-cleanup defaults to ON unless you tell it otherwise.

Easy!

The table level commands aren't any more complicated. Before we get started, please note that change tracking requires a primary key on the table you want to track. This is reasonable - you need some kind of unique identifier to tell you which row has changed.

With the PK requirement in mind, on we go!

Much as enabling change tracking at the database level uses a simple ALTER DATABASE command, enabling change tracking at the table level uses a simple ALTER TABLE command:
ALTER TABLE Demo_CT
ENABLE CHANGE_TRACKING
WITH (TRACK_COLUMNS_UPDATED = ON);

There's not a lot going on here either. You have three things to specify:
  1. The table you want to track changes on
  2. Enabling or (spoiler alert!) disabling change tracking
  3. Whether you want to track which column(s) were updated or not
As you may have guessed, if you want to disable change tracking, just change ENABLE to DISABLE and omit the WITH clause.

Speaking of that WITH, if you don't specify a setting for TRACK_COLUMNS_UPDATED, SQL Server will default it to OFF in order to reduce space requirements for change tracking.

While we're talking about TRACK_COLUMNS_UPDATED, it does just what it says on the tin. This is a good time to remind you that unlike other methods of tracking changes, all this does is tell you that a given record was updated and which columns were updated - it does not give you any state data. We'll cover this in more detail in a future post.

A word of warning: you can't modify the TRACK_COLUMNS_UPDATED setting without disabling and re-enabling change tracking on the table.

To see which tables have change tracking enabled, query the catalog view sys.change_tracking_tables:

As with sys.change_tracking_databases, the information is fairly basic. Kendra Little's fantastic article I referenced in my last post has a great script for this, too:
-- Kendra Little made this too; it's also posted at https://www.brentozar.com/archive/2014/06/performance-tuning-sql-server-change-tracking/
SELECT sc.name AS tracked_schema_name,
    so.name AS tracked_table_name,
    ctt.is_track_columns_updated_on,
    ctt.begin_version /*when CT was enabled, or table was truncated */,
    ctt.min_valid_version /*syncing applications should only expect data on or after this version */ ,
    ctt.cleanup_version /*cleanup may have removed data up to this version */
FROM sys.change_tracking_tables AS ctt
JOIN sys.objects AS so 
    ON ctt.[object_id]=so.[object_id]
JOIN sys.schemas AS sc 
    ON so.schema_id=sc.schema_id;

Of course, feel free to write your own if you need different information!

For our simple example, it gives us:

Joining on sys.schemas and sys.objects, Kendra's script yet again gives us human-readable results. I can't emphasize enough how important it is to script for readability.

As I mentioned before, disabling change tracking at the table level is equally simple:
ALTER TABLE Demo_CT
DISABLE CHANGE_TRACKING;

Don't forget - if you wish to disable change tracking at the database level, you'll need to disable change tracking for all tables first.

Thank you for continuing to read this series! Please join me next time when I show you around working with change tracking!

Sunday, August 11, 2019

Tracking Data Changes - Change Tracking, Part 1

Let's begin our adventure into tracking data changes in SQL Server with change tracking. Change tracking is a simple tool in SQL Server that does exactly what it says on the tin: tracks changes.

Microsoft provides excellent documentation on change tracking here. Microsoft has done a good job of retaining a lot of documentation for versions of SQL Server prior to 2014, but there are some gaps. Change tracking happily hasn't changed much since it was introduced in SQL Server 2008, so most of this information is valid for SQL Server 2012, 2008R2, and 2008.

SQL Server 2014 SP2 does introduce a documented stored procedure that you can invoke to clean up the internal change table (and don't worry, we'll cover this in due course!), but that's about it as far as documented functional changes go (note: if any of you hardcore internals folks know of under-the-hood changes between versions, even if they work the same on the surface, please let me know! I love that kind of detail!). That said, there's always the chance that future versions will introduce differences, so keep that in mind.

Anyway, moving forward! Here is a quick overview of change tracking:

  • The change source is an in-memory rowstore, flushed to disk with each checkpoint
  • It answers two questions:
    • Has a row changed?
    • What rows have changed?
  • The only values from the source table stored in the change table are the primary key values for the changed rows
  • You have to use the built-in functions to get at the change data
  • The database you want to track changes in must have a compatibility level of at least 90
    • Filed under 'ways the database engine will let you score an own goal', SQL Server will still let you enable change tracking on a database if the compatibility level is lower than 90, but the functions used to retrieve change data will give you an error

Change tracking is a fairly simple tool to set up and use. In order to enable change tracking, you must first enable it at the database level via an ALTER DATABASE statement, and then at the table level via an ALTER TABLE statement.

Let's begin at the database level. The ALTER DATABASE statement looks something like...

There aren't a lot of parameters here. You can set change tracking on or off, you can specify your retention period, and you can specify whether to enable auto-cleanup or not.

For the retention period, you have the choice of DAYS, HOURS, or MINUTES. If you don't specify retention, SQL Server will configure change tracking with a retention period of 2 days. If you provide a number but don't specify the interval, SQL Server will default to days. The minimum retention period is 1 minute.

For auto-cleanup, if you specify OFF SQL Server will not automatically clean up change tracking data. If you don't specify auto-cleanup, SQL Server will default to ON. Unless you want to be responsible for cleaning up after change tracking, ON is your best bet.

To see which databases have change tracking enabled, and to look at their configuration, query the sys.change_tracking_databases catalog view:



It covers the basics, but it could be more informative. Kendra Little has a nicer script:

For our simple example, it gives us:

By joining on sys.databases to provide the database name and omitting the retention_period_units in favor of the description, Kendra's script gives us much more readable results. You'll notice that this script omits the max_cleanup_version column. We'll come back to that, and when it's useful, in a future post.

As an aside, while it can sometimes be a pain to script for readability, it almost always pays off in the long run.

Disabling change tracking is similarly simple. As with enabling change tracking, you disable it with a very straightforward ALTER DATABASE command:

You will need to disable change tracking on all tables before disabling it at the database level, but as we haven't looked at the table level commands yet we'll just pretend we've already done that.

You can change the database-level settings for change tracking with a similar ALTER DATABASE script. Just omit the = ON or OFF:

You can modify either or both the retention period and the auto-cleanup setting.

That wraps it up for this post! Join me next time when we look at enabling change tracking at the table level!

Sunday, July 14, 2019

The best laid plans etc.

Hi there! I apologize for this blog going quiet again. Major project at work, all that, and this blog slipped down lower on my list of priorities outside of work, past spending time with my kids and my partner, exercise, rest, and all that.

Things are wrapping up, so look for a new post about tracking data changes soon!

Sunday, April 21, 2019

A brief note on version...

With Microsoft's move away from the Service Pack model, SQL Server has seen more frequent updates. This is generally a good thing, but it does pose a problem for those of us who write about it. Information gets out of date. New functionality comes in, old functionality is deprecated, documentation for older versions disappears, and the great cycle of technology continues. Many of us (like me!) write these blogs in our free time, to educate the community and ourselves, and it can be difficult to go back and maintain older posts to ensure they're up-to-date.

I have both Windows and Linux versions of SQL Server (at the time of this writing, my lab VMs are SQL Server 2017 CU14 on Windows 10 and Ubuntu Server 18.04 LTS; I have spun up a container or twelve in my time as well, and of course there's Azure SQL Database, and...), and I keep them up-to-date as new cumulative updates are released. I'll do my best to post version information as I write, but I probably won't go back and update old posts unless there's a really big change. Even then, I'll probably write a new post instead.

If you need historical version information, go directly to Microsoft SQL Server Versions List, do not pass GO, do not collect $200. Maybe toss 'em a couple bucks. You can also now find documentation for previous versions of SQL Server on the SQL Server previous versions documentation page.

Tracking Data Changes - Introduction

Unless you're managing a static archive database, the data in your databases will change. Odds are, at some point you're going to need to know about these changes for one reason or another. Perhaps you're using a data warehouse and need to retain history, or you need to keep track of slowly-changing dimensions. Perhaps you have an application that needs the latest data, and you want an easy way to feed it without querying the entire table for timestamps or update flags. Perhaps you have audit requirements and need to keep detailed change history.

Whatever the reasons, SQL Server has a few options to satisfy various use cases. Your options range from the 'roll up your sleeves and do it yourself' triggers, to Change Tracking, Change Data Capture, and Temporal Tables.  In this series, I'll show you these options and how to use them. We'll begin with Change Tracking, then move on to Change Data Capture, Temporal Tables, and conclude with triggers.

If you're using a platform other than SQL Server, please refer back to your RDBMS's documentation (or - let's be honest here - StackOverflow followed by your search engine of choice) for information about tracking data changes. Many platforms have similar functionality to handle this need, but if nothing else triggers are almost always a viable option, if a bit more manual work than we might like.

In SQL Server, some of the functionality depends on your version and edition. The following table provides a quick summary:

Tool Debut Edition
Change Tracking SQL Server 2008 All
Change Data Capture SQL Server 2008 Standard as of SQL Server 2016 SP1, Enterprise prior
Temporal Tables SQL Server 2016 All
Triggers Back in the dim mists All

Here is a (very) quick overview of the various options:
Change Tracking:
  • Uses an internal versioning mechanism to track changes
  • Just tells you that a change happened (it can tell you WHICH column was changed) - no history
  • Snapshot isolation, so be mindful of TempDB
  • Reasonably tolerant of schema changes
  • Requires a primary key
  • Automatically cleans up after itself
Change Data Capture:
  • Uses a log reader agent (hello sp_replcmds!) to capture change data
  • Can give you full change history
  • Does not play nicely with schema changes 
  • Allows two capture instances per table
  • Highly configurable - capture only columns you want
  • Automatically cleans up after itself
Temporal Tables:
  • Also requires a primary key
  • Also gives you the full change history
  • Uses period columns in the source table to feed the history table
  • Easy to query
  • In SQL Server 2017, introduced the capability to clean up after itself
Triggers:
  • Offer you amazing flexibility in designing your tracking
  • Let you capture metadata and stuff the other solutions may not allow
  • Require you to manage everything from soup to nuts
There is a lot more about each of these methods, of course, and in the course of this series I'll walk you through each in some detail. Please check back soon for an introduction to Change Tracking!

Monday, July 16, 2018

A quick 'save me from myself' thing in SSMS

Have you ever felt that 'down-elevator' feeling in the pit of your stomach when you realize you just executed a script against the wrong instance? I have.

It's of course possible to add name checks and all to only allow execution on a given instance, I'm going to go out on a limb here and guess that most people don't do that in most cases.

Happily, there's a quick and easy setting you can take advantage of to give yourself a visual indication that hey, maybe you're not where you thought you were: the 'Use custom color:' setting in the Connection Properties of the Connect to Server dialog box.

To get there, first click Options:



Check the box for 'Use custom color:', and hit 'Select...':




Pick your color, click 'OK' (I tend to use red for production, orange for QA/test, green for development, but go with whatever works best for you), and then click 'Connect':




From here on out, the status bar down at the bottom of the query window will be whatever color you've picked for that instance:




While various third-party tools offer similar functionality (often with a lot more flexibility and configurability), this is built-in to SSMS and has been around for quite a while. It's a great quick gut-check!