Why can’t we have column alias in ORDER BY?

You probably think something like “What! Of course we can!”. And so did I when I read the title for a post about this in a forum. Let us first establish that you can have a column alias in ORDER BY:

p.ProductID AS pid
,p.ListPrice AS pr
FROM Production.Product AS p
ORDER BY pr[/sql]

The problem was that a table name was also in the ORDER BY clause. And this is when it becomes weird. The problem isn’t the column alias, it is the table alias. Basically the desire is to write something like:

p.ProductID AS pid
,p.ListPrice AS pr
FROM Production.Product AS p

Above was apparently allowed in SQL Server 2005 2000 (I never did it myself, however). (Update: Erland Sommarskog caught my error, this was allowed in 2000, not 2005.) In order to understand why above is backwards, we need to consider the logical query flow. The first part of a query which is executed is the FROM clause (logically, the optimizer can do as it pleases as long as we get the same result). In above example, I renamed the table (for the rest of the query) from “Production.Product” to “p”. That means that we cannot refer to “Product” or “Production.Product” anymore in the query. However, SQL has imperfections and the topic of this blog is an attempt from MS to rectify such earlier imperfections (some of them), i.e., improving the product.

Continuing the logical query flow (WHERE, GROUP BY, HAVING), we towards the end find the SELECT clause. Here is where you can rename the columns (a.k.a. column alias) for the rest of the query. After the SELECT list, it is meaningless to talk about different tables (where the data originally comes from), all we have is the result of the query (so far). This is why, we can in the ORDER BY refer to a column alias (generally); ORDER BY comes logically after the SELECT list.

However, the original problem was that one also wanted to include a table alias, which shouldn’t be allowed. So, why is it allowed, in the first place – consider it makes no sense from a logical standpoint? For that we need to look as somethining weird, which is allowed (both in T-SQL and ANSI SQL) and that can be somewhat useful:

ProductID AS pid
,ListPrice AS pr
FROM Production.Product
ORDER BY ModifiedDate

We don’t need to return something to the client app in order to sort over it. Sure, that can be useful, but from a logical perspective, this is spooky. We sort over something which doesn’t exist! So, the imperfections of the SQL language is the reason things starts to break down. Basically, what we do is to sort over something as if we would have included it in the SELECT list. We are saying, in the ORDER BY clause something like

“Hey, dbms! Can you please pretend we included Product.ModifiedDate in the SELECT list so we can sort over it. Bur don’t return it to the client app, I don’t want to see that column.”

Where it really breaks down is when we combine table and column alias. Let me repeat the problematic query:

p.ProductID AS pid
,p.ListPrice AS pr
FROM Production.Product AS p

The problem is that it is meaningless to specify a table alias here, the column does exist in the SELECT list. The only reason we can refer to something table-something at all, is to let the DBMS pretend we included something additional from the base tables, so we could sort over it. Such as (allowed):

ProductID AS pid
,ListPrice AS pr
FROM Production.Product AS p
ORDER BY p.ModifiedDate

One could of course argue that below would be more logical (not allowed):

ProductID AS pid
,ListPrice AS pr
FROM Production.Product AS p
ORDER BY Product.ModifiedDate

Arguing that one form of two above is more logical than the other is pretty irrelevant, houever. Both are weird constructs, and the language is what it is. Anyone feel like checking the ANSI SQL standard which of above two (first, second or both) are allowed/preferred? No me, I’m off to grab some breakfast!


Watch out for old stuff

No, I’m not referring to me, I’m referring to options, config values and such which should really have been removed from the product a long time ago. I was reading a recent blog from Jonathan Kehayias and realize how much old stuff are still visible one way or the other in the product. There are of course commands which has more modern replacements (manage logins and users, attaching databases, changing SID for user etc), but keeing the old command for a few version is common sense. I’m thinking more of stuff where the functionality it gone ages, or even decades, ago, so why keep the knob for it?

Referring to Jonathan’s article, the LoginMode property has in SMO and policies four possible values.

  • “Unknown” maps to registry value 3 or higher which is clearly rubbish (somebody manually changed this to something weird).
  • “Mixed” maps to value 2 in the registry, Known as “SQL Server and Windows”, or “Mixed mode”.
  • “Integrated” maps to value 1 in the registry. Known as “Windows only”.
  • “Normal” maps to values 0 in the registry. But what is this, you think? There are only two values, right?

So, we have something strange exposed in SMO and policies, and it is even referred to as “normal”. This is one example if heritage which should have been cleaned up a long long time ago. We have to go all the way back to SQL Server 6.5 in order to explain this.

  • “Normal” (more commonly known as “Standard” back in the days) is more like what “Mixed” is today.
  • “Mixed” meant something slightly different back then compared to what it means today. We need to recall that back then, we didn’t have Windows authentication like how it is done today. We could map Windows account to SQL Server account using a tool, and this mapping could be used to do what we call today a “Windows authentication” or “trusted connection”. This setting meant that SQL Server could attempt a trusted connection based on the login name you specify, even if you in the connection string didn’t ask for a trusted connection. Confusing, huh? No wonder they got rid of the distinction bewteen “Normal” and “Mixed”.

As of 7.0, the only registry values used are 1 (Integrated) and 2 (Mixed). So why do SMO still expose 0 (Normal)?

We have plenty of other such examples. Look in the registry and you can see mapping for _, # and $. These were used in the mapping tool, so you could map a Windows login, which apparently has a backslash in the name and SQL Server would translate the backslash to some other character (_ by default), allowed as a SQL Server login name. Why do these still exist in the registry?
Or “Default login”? There no such thing as a default login in SQL Server anymore (there is something like a guest *user* but that is a different things from a login.
Or how about setting the master database to full recovery model? That is absurd, you still can’t do log backup and it still behaves like in simple recovery.

OK, enough of the rant. Back to meaningful work.

Do you remember Database Gateways?

I replied to a question in a forum today about intercepting and changing a SQL query, before it reaches SQL Server. I started typing a reply, and realized that… wow… “Does this take me on a trip down memory lane, or what!”. So, just for fun, here is what I wrote (slightly modified):

There used to be products on the market, called “Database Gateways”, something that looks like one type of DBMS but actually accepts the SQL queries and submits them to some other type of DBMS. Basically an “in between” piece of software. For instance I used to work with such a product, which looked like a SQL Server but submitted the queries to IBM’s mainframe DB2.

That market pretty much disappeared around when ODBC was born. Yes, we are talking ancient here, around 1992.

There was a leftover from this time period in SQL Server, called “Open Data Services”, which was basically an API allowing you to write something that looks like a SQL Server. There was even a sample app which allowed you to intercept the SQL queries and then re-submit them to a SQL Server.

This API was in turn transformed over time to what became known as “extended stored procedures”, which in turn was deprecated with SQL Server 2005 and the SQL CLR support.

Rebuilding system databases in 2008 R2

UPDATE: Apparently, a workaround is to use SETUP.EXE from install media (and not the locally installed SETUP.EXE). See the connect item mentioned below.

All my attempts so far to rebuild the system databases in 2008 R2 has failed. I first tried to run setup from below path:
C:\Program Files\Microsoft SQL Server\100\Setup Bootstrap\Release

But above turns out to be the 2008 setup program, not 2008R2 setup; even though I have no 2008 instanced installed (I have only R2 instances installed). Apparently, the 2008 setup program does a version check of the instance to be rebuilt and since it is > 10.50.0, the rebuild fails. Books Online for R2 the section about rebuilding system databases, has the above path, which obviously is incorret.

So, this lead me to think that the path to use is infact:
C:\Program Files\Microsoft SQL Server\100\Setup Bootstrap\SQLServer2008R2

Above does indeed contain the R2 version of the setup program. The problem now is that it refuses to rebuild for some other reason. I opened a question in the MSDN forum, and Xiao Min Tan opened a connect entry for this: https://connect.microsoft.com/SQLServer/feedback/details/564905/rebuilding-system-databases-for-sql-server-2008-r2.

Meanwhile, waiting for a resolution for above, I strongly recommend you to do both file-level backup as well as SQL Server backup of your system databases, from a good known clean state of them. Remember to do file-level backup of the database files, in case your instance won’t start.

(Here’s a prior blog post about rebuilding system database, btw: http://sqlblog.com/blogs/tibor_karaszi/archive/2008/08/29/rebuild-master-in-sql-server-2008.aspx

Are non-modified FILESTREAM files excluded from DIFFERENTIAL backups?

Short answer seems to be “yes”.

I got this from a forum post today, so I thought I’d test it out. Basically, the discussion is whether we can somehow cut down backup sizes for filestream data (assumption is that filestream data isn’t modified very frequently). I’ve seen this a few times now, and often suggestions arises to somehow exclude the filestream data and fo file level backup of those files. But that would more or less leaves us with the problem of the “old” solution: potential inconsistency.

So I thought that perhaps diff backups can be helpful here? I.e., perhaps SQL Server might be smart enough to in a diff backup exclude the filestream files that were already in the most prior database backup? Sure seems so, according to below script. (I’m on the road now, so scrips could be neater…):

USE master

NAME fstest_data,
FILENAME N'C:\DemoDatabases\DbFiles\a\fstest_data.mdf'),
NAME FileStreamTestDBDocuments,
FILENAME N'C:\DemoDatabases\DbFiles\a\Documents')
NAME 'fstest_log',
FILENAME N'C:\DemoDatabases\DbFiles\a\fstest_log.ldf');

--4.08 MB

--3.08 MB

CREATE TABLE fstest..t (
,Name VARCHAR (25)

INSERT INTO fstest..t
GO 10

--13 MB



Mirroring: what happens if principal loses contact with both mirror and wittness?

Imagine a database mirroring setup where you have two LANs with a WAN link in between. Let’s call them site A and site B. Say that principal is currently in site A, and both mirror and witness are in site B. I.e., we are running database mirroring with a witness, and assuming we are running safety FULL (synchronous), we have auto-fail over.

Now, what is really fail over when it comes to mirroring? the simple answer is that the mirror will finish the recovery process (UNDO) and make the database available. So, what if the mirror and witness becomes isolated from the principal? In out setup, this would happen is site A is disconnected from site B. Well, the fail-over protocol is as follows:

If the mirror cannot contact the principal, it asks the witness if it has contact with the principal. If the witness responds and says it doesn’t have contact with the principal, then they consider this a quorum. They are in majority and fail over occurs. I.e., the mirror will make the database available (and essentially becomes the principal).

But, think about the poor principal now. It is isolated and knows nothing about what happens at the other end. The only logical thing to do is to make the database inaccessible. If that didn’t happen, then we would have a case where both sites had the database available, and we wouldn’t want that!

This is nothing new, and it is clearly stated in Ron Talmage’s great whitepaper on mirroring: http://technet.microsoft.com/en-us/library/cc917680.aspx.

This is easy to test. Setup mirroring (you can even do it between three instances in the same OS). Make sure you are running safety full and have a wittness. Now you check what Windows process ID the principal has:


Based on above, you now use Task Manage to kill the *other* two SQL Servers. Make sure that SQL Server Agent isn’t running for those two, else agent will start those SQL Servers again. Management Studio will now show the database as “In Recovery”, and if you try to access it, for instance from a query window, you get below error:

Msg 955, Level 14, State 1, Line 1
Database Adventureworks2008 is enabled for Database Mirroring,
but the database lacks quorum: the database cannot be opened.
Check the partner and witness connections if configured.

How selective do we need to be for an index to be used?

You know the answer already: It depends. But I often see some percentage value quoted and the point of this post is to show that there is no such percentage value.

To get the most out of this blog post, you should understand the basic structure for an index, i.e. how the b+ tree look like. You should also understand the difference between a clustered and a non-clustered index. In essence, you should be able to visualize these structures and searches through them as you read the text. If you find that difficult, draw a few versions on a piece of paper and “navigate” through them by tracing through them with a pen or your finger. After a while, you will do this in your mind. For instance, check out the sections under this.

I’m not saying that we shouldn’t consider selectivity when designing indexes – of course we should! I’m not saying that one shouldn’t have some vague “feeling” about how much data to be return when making such decisions. What I will prove is that there is in reality no set percentage that the optimizer uses. The comment we usually see is something like:

“If we return more than 5% of the rows, then an index will not be used.”

Where did that 5% number came from? I can assure you that this is not some hard-wired number in the optimizer (except for an edge-case, see below). The optimizer aims at running the query with as low cost as possible. Considering the data access part (think WHERE clause and the condition), this is pretty much about reading as few pages as possible (few page-accesses).

Just to cut down a bit on the thread that might follow these types of blogs (“Hey, when I do this, your observations doesn’t match, your blog post is incorrect!”), let us first consider some special cases:

Clustered index
The clustered index *is* the data. If the search condition (SARG) is SEEKable, then SQL Server will obviously seek through a clustered index instead of scan it. Anything else would be stupid. There can be *other* search conditions that are more efficient, but we are considering one search condition at-a-time.

Non-clustered index that covers the query
This is pretty much the same argument as for above. Since all data is in the index (“covers the query”), not seeking it would be stupid. Again, there can be cheaper alternatives for any of the other search conditions, but we are considering one condition at-a-time.

The value is not known to the optimizer
This is what happens when you have a TSQL variable you compare against. Something like “colname = @v”. The optimizer has no knowledge of the contents of this variable. Either it uses density (where applicable, like “=”), as stored in the statistics information of the index. Where not applicable (like “>”, “<“, “BETWEEN” etc), then the optimizer actually do use some hard-wired percentage value. This value can change between versions so give it a spin of you want to know what value you have for your version/build number. Note that a variable is not the same thing as a parameter. SQL Server sniffs parameters (parameter sniffing). Read this for elaboration: http://msdn.microsoft.com/en-us/library/ee343986.aspx.

The search expression is not seek-able
I hope you know this already, but just to point it out. In most cases, having some calculation at the column side will void the ability to seek through the index. This should ideally be known to all T-SQL developers: Never do calculations at the column side! So, things to avoid are like “colname * 23 > 45” or “SOMEFUNCTION(colname) = 44”.

Hopefully by now we all understand that there are always special cases and exceptions. The more of Kalen’s books you have read, the more you understand this. What we are discussing here is the typical situation. OK? Fine. So, “Why is there no percentage value that the optimizer uses?”, you ask. Because the value will differ. In short, SQL Server wants to read as few pages as possible. In the most simple example, the alternative to an index seek is a table scan. So we will use this as basis for your discussion. There can be other alternatives to the table scan (using some other index for some other condition), but that doesn’t change the principal “it depends” concept.

In essence, it is all about the alternative. As I said, our example wil use a table scan as alternative. A table scan (or clustered index scan if it is a clustered table) means that SQL Server will look at every page and see what rows satisfies the search condition on each page.

My example has two different tables, both with 100,000 rows. These tables both have an integer column with consecutive increasing unique values, which also has a non-clustered index. I will see how selective I need to be when searching on this column in order for an index search to be done, compared to a table scan. I.e, find this percentage cut-off value.

The fewrows table only fit one row per data page. This means 100,000 data pages. My tests show that the cutoff for fewrows is about 31,000 rows. I.e., we need to be more selective than 31% for the index to be used.

The manyrows table fit 384 rows per page. This means 260 pages. My tests show that the cutoff for fewrows is about 155 rows. I.e., we need to be more selective than 0.16% for the index to be used.

You might end up with different exact numbers, depending on what you have in the statistics, the build number of your SQL Server etc. But what you will see that a similar pattern. A huge difference between the two.

It is all about the alternative
If I look at my indexes using sys.dm_db_index_physical_stats, I will see that the non-clustered index on the int column for the two tables are exactly the same (same number of pages in the index, etc). So, two indexes with the same characteristics have very different cut-off values. How can that be? It is because the alternative differs. The alternative for this example is a table scan. For the bigrows table, the alternative means reading 100,000 pages. But for the smallrows table, the alternative means reading only 260 pages. There can of course be other alternatives, like using some other index for some other search condition. This is, in the end, why we don’t have a set percentage value: it is all about the alternative!

The typical cases will of course fall somewhere between my more extreme examples. But my examples show that there is no set percentage value used by the optimizer. I showed that for my test, the percentage value can be as low as 0.15% or as high as 31%. What matter is the alternative!


USE tempdb



INSERT INTO manyrows
FROM sys.columns AS asys.columns AS b

FROM sys.columns AS asys.columns AS b

CREATE INDEX ON manyrows (c2)
CREATE INDEX ON fewrows (c2)

--Number of pages etc:
EXEC sp_indexinfo 'manyrows'
-- Data: 265 pages (2.07 MB)
-- Index x: 187 pages (1.46 MB)

EXEC sp_indexinfo 'fewrows'
-- Data: 100385 pages (784 MB)
-- Index x: 187 pages (1.46 MB)

FROM sys.dm_db_index_physical_stats(DB_ID(), NULL, NULL, NULL, 'DETAILED')

--Run below with showplan:
SELECT FROM manyrows
--155 = ix search, 156 = ts
--Cut-off is 0.16%

--31000 = ix search, 32000 = ts
--Cut-off is 31%

Restore database to the point of disaster

This is really basic, but so often overlooked and misunderstood. Basically, we have a database, and something goes south. Can we restore all the way up to that point? I.e., even if the last backup (db or log) is earlier than the disaster?
Yes, of course we can (unless for more extreme cases, read on), but many don’t realize/do that, for some strange reason.

This blog post was inspired from a thread in the MSDN forums, which exposed just this misunderstanding. Basically the scenario was that they do db backup and only log backup once a day. Now, doing log backup that infrequent is of course a bit weird, but that is beside the point. The point is that you can recover all the way up to the point of disaster. Of course, it depends on what the disaster is (don’t expect too much if the planet blows up, for instance).

Since “log backup only once a day” was mentioned, I will first elaborate a bit on frequency for database vs log backups. For the sake of discussion, say we do both db and log backup once a day. You say:
“What? Both db backup and log backup once a day – why would anybody do that way? Wouldn’t one do log backup more frequently than db backup?”
Yes, of course (but I actually see such weird implementations from time to time). But again, that doesn’t change the topic at hand, but I will first elaborate on this; just so we don’t see blurring comments later arguing this irrelevant argument.

So, lets first sort out two different cases:

A) Log backup before the db backup
1: db backup

2: log backup
3: db backup
Here we will use backup 3 when we later will restore.

B) Db backup before log backup
1: db backup

2: db backup
3: log backup
Here we will use backup 2 and 3 when we later will restore.

You see that A) and B) are really the same thing? What is relevant is that we have all log records available (in ldf file/log backups) since the db backup we chose to use as starting point for the restore. Actually, for A), we could might as well use backup 1 and 2 (and skip 3)!

“Hang on”, you say, “we’re not done yet. What about the modifications since the last log backup! Gotcha!”
No worries, this is where it gets interesting, and below is really the heart of the topic. Clearly, we need to get the log records out of the ldf file into a log backup (file). If we can do that, then we will call this backup number 4, and use as the last backup for our restore. After doing that restore, we have no data loss!

So, how do we produce a log backup after a disaster?
It depends on the disaster! Let’s discuss a few scenarios:

a) Planet Earth blows up.
No can do. I doubt that anyone of you has mirrored data centers on Moon or Mars; and also people stationed off-Earth for these situations. Of course, I’m being silly. But my point is that you can always have a disaster such that you can’t produce that last log backup. No matter how much you mirror: if the disaster takes out all mirrors, then you are toast. Remember that when you talk SLA’s. That fact is not popular, but it can’t be argued. It is all about limiting the risk exposure – not eliminating it. Anybody who believes we can eliminate risk exposure is dreaming. Agreed? Good. Let’s move on to (hopefully) more realistic scenarios:

b) Something happens with the data file(s), lost, corrupt or so. Ldf file is still there.
This is the easy case, but so often overlooked. What you do now is to backup the log of the damaged database, using the NO_TRUNCATE option. Something like:
BACKUP LOG dbname TO DISK = ‘C:\dbname.trn’ WITH NO_TRUNCATE
Yes, it really is that simple. Then restore backups from above, including this last log backup. Don’t believe me? Test it.

  1. Create database and table
  2. Insert some data
  3. Do db backup (1)
  4. Insert some more data
  5. Do log backup (2)
  6. Insert some more data
  7. Stop SQL Server
  8. Delete mdf file
  9. Start SQL Server
  10. Do log backup using NO_TRUNCATE (3)
  11. Restore 1, 2 and 3.

c) Something happens with the database. Ldf file is NOT still there.
Clearly, if the ldf file is really gone, we can’t do a log backup – how much as we might want to. Remember the old days, when redundancy for disks (RAID) wasn’t as common as today? “If there’s anywhere you want redundancy, it is for the transaction log files!”

d) Something happens with the data file(s), lost, corrupt or so. Ldf file is still there. The installation is toast – we can’t start SQL Server.
This seems a bit more nerve-wracking, right? Not to worry, just do the right steps and you will be fine. You probably ask now:

“But how can we backup the transaction log when our SQL Server won’t start?”

That is a good question. You need to get that ldf file to a healthy SQL Server, and make SQL Server believe this is the ldf file for a broken database on that instance. It is not really complicated. Just use a dummy database on that SQL Server as intermediate – to get the right meta-data into that SQL Server, so in turn it will allow you to produce this last log backup. I will show just that:

I have two instances on my machine (named “a” and “b”). I will create and damage a database on instance a, and then produce a log backup for that orphaned ldf file a different instance, b. I will pretend these are on two different machines, using separate folders for the database files “C:\a” and “C:\b”. Here’s the T-SQL, starting with instance a:

x.dbo.t(c1 INT IDENTITY)

–Stop SQL Server and delete below file
–Start SQL Server

–Oops, damaged database…:

–Stop SQL Server, pretend installation is toast

Do we agree that we have a damaged database, and there has been done modifications since the last log backup? Ok, fine. We now pretend that SQL Server instance “a” doesn’t start anymore. So, I will try to produce a log backup from that ldf file on instance “b”:


–Stop SQL Server and delete below files

–Copy the C:\a\x_log.ldf to C:\b\x2_log.ldf

–Start SQL Server

–Produce our last log backup:

–Restore the database, up to last transaction.
–Investigate logical file names for MOVE options first:

,MOVE ‘x’ TO ‘C:\b\x.mdf’
,MOVE ‘x_log’ TO ‘C:\b\x_log.ldf’


–Data there?
–See? That wasn’t so difficult.

Note how I even named the dummy database differently on instance b, with different physical file names and different file sizes (all compared to what we had on instance a). Typically, you will use same database name and same filename, but I want to show that we don’t really have to know a whole lot about the damaged database in order to produce a log backup from the ldf file!

Case closed.


Are log records removed from ldf file for rollbacks?

Seems like a simple enough question, right? This question (but more targeted, read on) was raised in an MCT forum. While the discussion was on-going and and I tried to come up with answers, I realized that this question are really several questions. First, what is a rollback? I can see three different types of rollbacks (there might be more, of course):

  1. Regular rollback, as in ROLLBACK TRAN (or lost/terminated connection)
  2. Rollback done by restore recovery. I.e., end-time of backup included some transaciton which wasn’t committed and you restore using RECOVERY, so SQL Server need to rollback this transaction (UNDO).
  3. Rollback done by crash (restart) recovery. I.e. SQL Server was shut down with some open transaction.

I’m going to try to show whether log records are removed or still present for these three types of rollback situations. I will use the fn_dblog function. This isn’t documented or supported, but search the internet and you will find how to use it. The result isn’t documented either, of course, so we have to guess a bit what the various values mean…

The TSQL script has some common parts (the same part executed repeatedly, once for each test):

  1. Create a database
  2. Make sure it is in full recovery
  3. Do a database backup
  4. Create a table (myTable)
  5. Insert a row into myTable. This last operation generates 5 log records for the myTable table: one for the PFS page, two for IAM pages, one format page for the heap and the last one is a LOP_INSERT_ROWS.
  6. Start a transaction
  7. Insert one more row into myTable. We now have one more log record for myTable (LOP_INSERT_ROWS). Looking at the transaction id for this last insert, we see two log records (one LOP_BEGIN_XACT and the LOP_INSERT_ROWS). Note that this transaction is now open!

Here is the above mentioned first part of the script:

USE master
TO DISK = ‘C:\x.bak’ WITH INIT


FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–5 rows


FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–6 rows

WHERE [Transaction ID] (SELECT TOP(1[Transaction ID] 
FROM fn_dblog(NULL, NULL)
WHERE AllocUnitName LIKE ‘%myTable%’
–2 rows

Now, on to the different cases: 

1. Regular rollback, as in ROLLBACK TRAN.
It seems reasonable to me that SQL Server will just add some “rollback transaction” log record here. So, let’s try that (continuing on above first part)… We now have 7 log records for myTable, with an added LOP_DELETE_ROWS which undo the previously insert. And for our transaction ID, we have 4 rows, with added two rows being LOP_DELETE_ROWS (compensation log record) and a LOP_ABORT_XACT.


FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–7 rows

WHERE [Transaction ID] (SELECT TOP(1[Transaction ID] 
FROM fn_dblog(NULL, NULL)
WHERE AllocUnitName LIKE ‘%myTable%’
–4 rows

2. Restore recovery
Now, what does this really mean? Restoring a transaction log backup is a good example. SQL Server read log records from the transaction log backup file and writes them into the LDF file. This is what we call the “data copy” phase. Then SQL Server performs REDO (a.k.a. roll forward). And finally, SQL Server performs UNDO (roll back), assuming we don’t do the restore using the NORECOVERY option. Restoring from a database backup isn’t any differene except the log records are of course read from the database backup file.

Here it seems likely that SQL Server will wipe more or less anything from the LDF file as soon as the database is restored and brought on-line. Why? The log in this newly restored database can’t serve as a starting point for a restore operation for this database. You first need a database backup. So, no use hanging onto log records either! Let’s see if we can verify that:

–Restore recovery

–Do this from different connection
–We now have open transaction in database!

–Perform RESTORE

–Rollback was done, and database berought online. 
–What log records do we have?

SELECT FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–0 rows

–Use the transaction ID from earlier SELECT
WHERE [Transaction ID] ‘0000:00000203’
–0 rows

So, SQL Server will remove user-defined stuff from LDF file after restore recovery was performed. Makes sense.

3. Crash recovery (a.k.a. restart or startup recovery)
I couldn’t really guess here. So, lets give it a spin immediately and see:

–Crash recovery

–Do below from separate connection

— startup SQL Server and examine the log records:

SELECT FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–5 rows

–Use the transaction ID from earlier SELECT
WHERE [Transaction ID] ‘0000:00000204’
–0 rows

Hmm, so SQL Server removes the log record after the rollback was performed. Makes sense.

But this got me thinking some more. How can this be done… physically? SQL Server would just “reverse” the head of the log a bit. But what if we have later transactions for other connections, which has been committed? SQL Server can’t ignore those, of course. These need to be kept in the LDF file for subsequent log backups. OTOH, I doubt that SQL Server will somehow physically delete things “in the middle” of an ldf file. Time for yet another test:

–Crash recovery, part two

–Do below from separate connection
–First something which creates more recent log records

— startup SQL Server and examine the log records:

SELECT FROM fn_dblog(NULL, NULL) WHERE AllocUnitName LIKE ‘%myTable%’
–7 rows

–Use the transaction ID from earlier SELECT
WHERE [Transaction ID] (SELECT TOP(1[Transaction ID] 
FROM fn_dblog(NULL, NULL)
WHERE AllocUnitName LIKE ‘%myTable%’
–4 rows

Now the log records for our rolled back transaction are still there! So, just as when we did a regular rollback, SQL Server inserted a LOP_DELETE_ROWS to reflect the undo of the INSERT, and then a LOP_ABORT_XACT.

Isn’t it beautiful when it all makes sense? Here are my conclusions, whether log records are kept or removed from transaction log file (LDF) for various types of rollback scenarios:

  • Regular rollback. Log records are not removed. Compensation log records are logged, reflecting undo of the modifications, and then an LOP_ABORT_XACT is logged.
  • Restore recovery. Log records are removed.
  • Crash recovery. It depdends. If you have a transaction which is at the very head of the log, then those log records can be removed. If there are other, subsequent committed transactions, then compensation log records are logged, reflecting undo of the modifications, and then a LOP_ABORT_XACT is logged.