SQL Server 2016 still uses mixed_extent_allocation in TempDB

Question

Today I was troubleshooting a problem on TempDB on SQL Server 2016 SP2 using extended events and added the mixed_extent_allocation and transaction_log event to my trace.

I expected mixed_extent_allocation wouldn't show up in my results because following query returned 0 as result:

select is_mixed_page_allocation_on  
from sys.databases 
where database_id=2

But to my suprise this event showed up several times. This was preceded by a transaction_log event with SGAM as context and operation LOP_SET_BITS.

This made me curious and I checked the content of the first SGAM page of TempdB:

DBCC TRACEON(3604)
dbcc page(tempdb,1,3,3)
DBCC TRACEOff(3604)

This is a snippet of the result:

╔═════════════════════════════════════════════╗
║ (1:0)        - (1:176)      = NOT ALLOCATED ║
║ (1:184)      -              =     ALLOCATED ║
║ (1:192)      -              = NOT ALLOCATED ║
║ (1:200)      - (1:208)      =     ALLOCATED ║
║ (1:216)      - (1:256)      = NOT ALLOCATED ║
║ (1:264)      -              =     ALLOCATED ║
║ (1:272)      -              = NOT ALLOCATED ║
║ (1:280)      -              =     ALLOCATED ║
║ (1:288)      - (1:296)      = NOT ALLOCATED ║
║ (1:304)      -              =     ALLOCATED ║
║ (1:312)      -              = NOT ALLOCATED ║
║ (1:320)      -              =     ALLOCATED ║
║ (1:328)      - (1:336)      = NOT ALLOCATED ║
║ (1:344)      -              =     ALLOCATED ║
║ (1:352)      - (1:65528)    = NOT ALLOCATED ║
╚═════════════════════════════════════════════╝

This made me conclude that mixed extents are still used. I thought SQL Server 2016 only uses uniform extents (except for master, msdb and model). Or is my conclusion wrong?

Why am I still seeing the mixed_extent_allocation event popping up?

Answer 1

According to this post, IAM pages are still allocated from mixed extents, even in SQL Server 2016. So this will be the mixed_extent_allocation I saw in my trace.

The blog post by Paul Randal provides some more internals explanation. Thanks to Scott Hodgin for pointing me to the article.

I did soms more research about the content of de pages allocated in the first SGAM page and found out that SQL Server also uses mixed extents for system tables (object_id lower than 100).