Skip to Content

… wondered why SQL Server behaves to weird 😕

Have you ever asked yourself questions like:

  • why is my transaction log running full if I’m already using recovery model simple?
  • how often should I update the statistics of the database objects?
  • how often should I reorganize or rebuild tables and indexes?
  • why is the timestamp of the optimizer statistics for some objects not new if my Update_Tabstats job runs frequently?
  • why is so much data missing in some tables after I used repair_allow_data_loss to repair database inconsistencies?
  • why does DBCC CHECKDB or DBCC CHECKTABLE still find inconsistencies when I’ve already used repair_allow_data_loss?
  • why are my datafiles not growing the way I expect them to even though I’ve configured the files to autogrow?
  • why is my database occupying less space after I’ve archived so much data?
  • why is my table not occupying less space after I’ve deleted so many rows from it?
  • why is the result of my query not ordered anymore even if it always used to be?

Bad news first:

My experience says:

NO, to 99,99999 % what you see is NOT a bug. Instead, there’s simply a gap between what you think how it’s supposed to work and how Microsoft designed it to work.

… and now the good news:

YES, there IS a comprehensive explanation to your: but why?!?! questions and finally you get all the anwers at once 🙂

To clarify all these frequent misconceptions I released:

For those of you who don’t have access to SAP Notes and KBAs I once paste the current content of the note here ….

If you come across similar topic, let me know and I’ll try my best to cover them as well.

Regards,

Beate

———————————————————————————————————————————————————————-

1660220 – Microsoft SQL Server: Common misconceptions

———————————————————————————————————————————————————————-

Some widely accepted information about the Microsoft SQL Server database is shown not to be entirely correct upon closer examination. This Knowledge Base Article lists widespread incorrect assumptions about Microsoft SQL Server and explains why they are wrong.

  1. The Microsoft SQL Server Agent job Update_Tabstats updates the database statistics which are used by the database optimizer to calculate execution plans and therefore it is critical for performance if the job fails. 
  2. When using recover model simple the transaction log cannot run full. 
  3. The result set of database accesses is always ordered by the primary key even if no ORDER BY clause is used explicitly. 
  4. If database accesses hang for a long time, the problem is caused by a deadlock. 
  5. Updating Microsoft SQL Server database statistics manually (with a SQL Server Agent job, with a SQL Server Maintenance Plan or by other means) is part of maintenance and therefore required on a regular basis. 
  6. Reorganizing some or all database objects is a required maintenance task and should therefore be carried out on a regular basis. 
  7. DBCC CHECKDB and DBCC CHECKTABLE with option repair_allow_data_loss allow you to repair database inconsistencies and will not cause any data loss. 
  8. After archiving or deleting data from a table the table and its indexes will occupy less space in the database and the database itself will also occupy less space. 
  9. If the autogrow option is configured for all datafiles Microsoft SQL Server will grow all files in a balanced way.

1. The SQL Server Agent job Update_Tabstats updates the database statistics which are used by the database optimizer to calculate execution plans and therefore it is critical for performance if the job fails.

Job SAP CCMS_<sid>_<SID>_Update_Tabstats does not touch the optimizer statistics at all and has no influence on execution plans.
Instead, it is part of the database monitoring framework implemented and provided by SAP.
It is not natively included in a Microsoft SQL Server installation, but is developed and delivered by SAP and was introduced with Basis Release 7.00 SP12.

The job executes stored procedure sap_update_tabstats.
It collects meta information about database objects (e.g. keyfigures like the number of table rows, the reserved size of an object, the row modification counter, and many more) and stores them persistently in the database.
As SQL Server does not keep any history for such keyfigures, SAP collects and stores them with this job in order to make historical information about database objects available.

This allows to analyze how certain properties of database objects change over the time and serves as source of information for SAP DB monitoring transactions (e.g. DBACockpit, fastest growing tables, …).

Also see SAP Note 1178916 for more information.

2. When using recovery model simple the transaction log cannot run full.

The transaction log of a database consists of one or more files. You can decide for each file if you allow SQL Server to autogrow the file if required or not (autogrow on/off). With recovery model simple you ensure that SQL Server will truncate the log at each checkpoint – still, this doesn’t mean that the transaction log cannot run full. Imagine you have a very long running transaction and all transaction log space is consumed before the transaction reaches the point in time where it commits. In such a case the transaction log can run full even if you are using recovery model simple. To resolve you need to take a closer look at the transactions – is there a long running transaction which keeps SQL Server from truncating the log? Is it normal that this transaction takes so much time or is it caused by a wrong executiong plan, bad I/O or any other performance-degrading issue?

To understand this in detail you need to make yourself familiar with how transaction log truncation works, meaning: which parts of the transaction log are considered active and which are considered inactive when the truncation is carried out. Parts cannot be truncated as long as they are still active. For a detailed example please see SAP Note 421644.

3. The result set of database accesses is always ordered by the primary key even if no ORDER BY clause is used explicitly.

The result set is only ordered by the key if SQL Server uses the primary index for the access.
Since the database optimizer chooses the access path dynamically based on existing statistics, you cannot assume that SQL Server ALWAYS uses the primary index for certain accesses and therefore cannot rely on an ordered result.
If you require the result of a query to be ordered you must use an ORDER BY clause.

4. If database accesses hang for a long time, the problem is caused by a deadlock.

This assertion is incorrect.
Genuine deadlocks (in other words, the mutual blocking of several transactions) are quickly recognized by Microsoft SQL Server, and eliminated by canceling one of the blocking transactions with an SQL Error 1205 within seconds.
Database accesses that hang for a long time may have a wide variety of causes (blocking locks or suboptimal execution plans for example), but are not deadlocks.

A good starting point to analyze why certain actions hang is creating snapshots of the current situation with hangman. Refer to SAP Notes 948633, 541256 and 806342.
If you really encounter a deadlock (which becomes evident by the occurence of an SQL Error 1205), you can analyze it in more detail by following SAP Notes 111291 and 32129.

5. Updating Microsoft SQL Server database statistics manually, with a SQL Server Agent job or a SQL Server Maintenance Plan is required on a regular basis.

Updating statistics on a regular basis is important but with SQL Server there is no need to schedule this task explicitly!
As long as the autostats feature is properly enabled, Server automatically detects if statistics need to be updated and carries out this task for you.
The decision if statistics are considered out of date depends on several factors like the number of rows in the table and the number of rows modified since the last statistics update.

For details on the algorithm please see:

  To ensure that the automatic statistics update is enabled correctly, please refer to the configuration note for your SQL Server release: 

  • SQL Server 2000: SAP Note 327494 
  • SQL Server 2005: SAP Note 879941 
  • SQL Server 2008: SAP Note 1237682 
  • SQL Server 2012: SAP Note 1702408

Bottom line:

Don’t update optimizer statistics manually for any object (or the whole database) unless SAP explicitly asks you to do so. It will produce I/O load and will not have any benefit.

The only exception to this rule are tables which contain date information. To ensure proper statistics for those tables at any times you need to follow SAP Note 1558087 and schedule an update job for such tables.

6. Reorganizing some or all database objects is a required maintenance task and should therefore be carried out on a regular basis.

A bad overall performance or a bad performance of single database operations is often believed to be caused by the fragmentation of tables and indexes. As a solution, reorganization or rebuild appears to be the cure. This might apply for other relational database management systems, but for SQL Servern in most situations both, the assumption that the bad performance is caused by fragmentation, and trying to solve the problem by reorganizing or rebuilding database objects, are false.

For this reason, SAP explicitly recommends not to reorganize or rebuild any database objects on a regular basis. You should not even reorganize or rebuild objects as an attempt to solve a performance problem as long as it is not evident that fragmentation is the root cause of the problem (which it hardly ever is).
Please see SAP Note 159316 which explains this topic in more detail.

7. DBCC CHECKDB and DBCC CHECKTABLE with option repair_allow_data_loss allow you to repair database inconsistencies and will not cause any data loss.

The repair_allow_data_loss option is not a tool which can perform magic to retrieve back data from pages which are physically damaged or contain logically incorrect information. Instead, it does more or less exactly what its name says: it tries to retrieve as much data as possible and will discard as much data as required to return to a consistent version of the affected object(s).

It is important to understand that database inconsistencies are in almost all cases caused by malfunctions on lower layers (typically hardware or driver malfunctions). This means that due to a malfunction on these lower layers one or more database pages are damaged – meaning their content is not fully correct anymore to a certain extent. There are various types of database inconsistencies e.g. pages might not linked properly anymore, links between pages might be missing completely, pages from the allocation maps (GAM, IAM, SGAM) might contain incorrect data, pages might be damaged to an extent that they do not even have the physical structure of a SQL Server page anymore.

If you are very very lucky this affects a page which was cached for faster access in your main memory and the inconsistent page hasn’t yet been written back to the disk. This is what we call a transient inconsistency but unfortunately an inconsistency is hardly ever a transient one.

In most cases the inconsistent pages are in the database files or in the log files. This means the incorrect information is on disk and there is no proper version of the affected page(s) anymore. This should make it clear why you cannot simply “recover” from an inconsistency. An inconsistency is a damaged page – there is no way to make the database guess what the correct content of an inconsistent page would have been and to let the DB simply revert the page content back to the correct version.

In most cases you will have more than one inconsistency. In order to judge how bad the situation is you need very exhaustive knowledge of SQL Server to understand which kind of pages (e.g. index pages, data pages, leaf pages, allocation map pages) are affected and which impact this has.

Using repair_allow_data_loss in order to let the database discard everything that cannot be interpreted or read properly anymore is no solution – in most cases it will even make things worse and still there is no guarantee that this option will even be able to recreate a consistent version of the affected object(s) – despite accepting data loss. You might still have inconsistencies left afterwards as depending on how bad the situation is, it might not even be possible anymore to return to a physically consistent state. On the other hand and much more important: this leads to completely uncontrollable, unpredictable dataloss and there is no way to log or trace what is thrown away. You will have data loss and this will cause inconsistencies on SAP application level (usually SEVERE inconsistencies). For these reasons, SAP does not support the usage of repair_allow_data_loss. See SAP KBA 1704851 and SAP Note 142731 for further details.

8. After archiving or deleting data from a table the table and its indexes will occupy less space in the database and the database itself will also occupy less space.

There are different key figures which inform you about the space consumption of an object (reserved size, data size, index size, unused size). If you have a large table and you delete a large amount of data from the table (e.g. by reorganizing the entries from application level or by archiving) SQL Server will not release the freed space back to the data files. Instead, it will keep the freed space reserved for the object.

If you really have the need to release the space back to the datafiles and to then release it back to the filesystem, please refer to SAP KBA 1721843 for more details. If you are not urged to gain back the space on filesystem level, SAP recommends to simply leave the object as it is. SQL Server will reuse the freed space as soon as new entries are inserted into the table.

9. If the autogrow option is configured for all datafiles Microsoft SQL Server will grow all files in a balanced way.

The assumption that MS SQL Server grows files in a balanced way if autogrow is switched on for all files is a common misconception.
SQL Server uses a proportional filling algorithm to distribute new data over all existing datafiles. This is described in more detail in SAP Note 1238993 – even though the note explicitly mentions SQL Server Release 2008 it works the same way for all releases.

Briefly explained: if new data needs to be added to the database, SQL Server distributes the new data over all datafiles which still provide free space. SQL Server will not grow any file as long there is at least one datafile which still provides free space.
Even if autogrow is configured for all datafiles SQL Server will wait until ALL files are full – only if all datafiles are completely full and SQL Server needs to add new data it will autogrow files. If you are using SQL Server release >= 2008 and have set trace flac 1117, SQL Server will grow all existing datafiles with autogrow=on at an autogrow event. In any other case, SQL Server will grow a single file only! All new data will then go into this single grown file until this file is full again. Then, as soon as new data needs to be added again, SQL Server will repeat the previously explained procedure and will grow a single file only.

In order to allow MS SQL Server to use all files at any time, we strongly recommend you to make sure that you always have free space left in all existing datafiles. If some of your datafiles are currently full, extend them if your disk layout allows it.
For SQL Server Releases >= 2008 Microsoft provides trace flag 1117 which makes SQL Server grow all files instead of growing a single file only at an autogrow event. This reduces monitoring efforts to ensure proper data distribution. Please see note SAP Note 1238993 for more details and to learn how to set the trace flag.

To report this post you need to login first.

1 Comment

You must be Logged on to comment or reply to a post.

  1. Edwin Sarmiento

    Thanks for posting this great list, Beate. I’d like to add this reference from Paul Randal, CEO of SQLSkills.com and former Program Manager on the SQL Server Storage Engine Team. He wrote a series of blog posts on the different myths regarding SQL Server and compiled them in a PDF format.

    A couple of comments on the items described,

    item 3The result set is only ordered by the key if it is defined as the CLUSTERED INDEX. By default, the primary key on a column is automatically created as a unique clustered index unless otherwise specified. The clustered index contains the actual data arranged in order of the defined indexed column.

    item 4 – A great way to find out what is causing SQL Server to wait is to use the waits and queues methodology in performing performance troubleshooting.

    item 5 – A proper understanding of the data volume and data modification behavior is key to deciding whether or not manually updating statistics would be helpful. Statistics only get updated automatically if the leading column defined in the statistics object has changed by more than 500 + 20% of the number of rows in the table. Imagine a 2-million-row table where you need at least 400,500 modifications in order for the statistics update to kick in. If modifications don’t hit that rule, statistics won’t get updated and the query optimizer will use an outdated statistics to generate a suboptimal query plan.

    item 6 – Fragmentation is caused by data modification and definitely affects query performance. Imagine a phone book ordered by Lastname. If a new entry needs to be added in the phone book, we need to check the value of the Lastname, check to see if we have free space on the page where we can insert the record. If we don’t, we need to create a new page before we can add the new page. In SQL Server terms, this is called a page split. In order to maintain the links between the old page and the new page, SQL Server has to write pointers to both pages and maintains them. Wouldn’t it be easier to find records in contiguous pages as compared to noncontiguous pages that can onlybe navigated by knowing the pointers that lead to them? That is how fragmentation happens. If the database is experiencing high transaction volumes, fragmentation is but a natural side effect and this is why regular maintenance is recommended by Microsoft, even on SAP systems.

    I do, however, agree that before even performing an index rebuild or reorganization, make sure that the objects are indeed fragmentation. Performance issues can be caused by a ton of different other factors – memory, CPU, network, IO, etc.

    (0) 

Leave a Reply