2021-11 When is an NPI not an NPI?

This month, I wish to discuss everything about indexes but focusing on Non-Partitioned Index’s (from now on NPIs). I must confess that the real name of an NPI is a Non-Partitioned Secondary Index but I really dislike that as a term…

In the Beginning was the Index

Well, actually, it was the data, but then we needed a quick way of accessing the data and so the index was born. In Db2, these are b-tree based indexes with a few system pages and then a hierarchy of non-leaf pages starting with the root page, which then branches down in levels until it finally hits the leaf page – and these contain the direct pointers to the data. Very fast and just a few getpage I/Os to read the non-leaf/leaf pages in.

Hope Springs Eternal

The hope is that all the system stuff, the root page, and quite a few other non-leaf pages will be hanging around in the bufferpool and so no real I/O is required, but you all know the real world? It never is like that!

Bigger and Bigger

Tables and tablespaces got bigger and bigger over time, so then the Partitioned Tablespace was born. At first it required a partitioning index (PI), but nowadays you should all be using table based partitioning.
Brief aside: Recently, there was a mega outage caused by a really nasty bug in Db2 that brought everything crashing down (PH41947). Root cause was the use of index-based partitioning – still! Do IBM test all these old things on all levels of code? Nope. Time to get my Migration HealthCheck freeware and migrate away from index-based partitioning as fast as you can!

A Star was Born

Well ok, not really a star but the PI started life a little bit dazed and confused, and the formula for calculating “how big can my PI can get before it goes *boom*” is also impressive! (Check out my Know Your Limits presentation for the formulae!)

And Then Came the NPI

Now this PI was naturally unique, but we all needed a way to check for other data and possibly non-unique column combinations. So the non-partitioned index (NPI) was born. This is an index over the *complete* partition range and was normally heavily used by application SQL.

Pain points began

First major problem was the REORG of a single partition. It did the Tablepart (TP) data and it did the Index Part (IP), but it also had to update all the pointers within the NPI. This used to be terribly slow. At one shop, when reorging 22 partitions of a 254 partition TS it took 26 hours at the TP level and only 18 hours at the complete TS level. This is when we added to our RealTime DBAExpert utility generation software the ability to “roll up” from a TP reorg to a TS reorg based on number of partitions being REORGed, percentage of partitions being REORGed, or whether or not at least one NPI existed.

A Piece of Cake?

The other major problem with NPIs is that they can have a PIECESIZE, which is how large a Linear Page Set (LPS) can get before a new one is created. You can have from 1 to 4096 pieces for a single NPI, and the calculation to work out how many pieces you are allowed to have is another brain-numbingly complex one – which is also in my limits presentation.

Some sample SQLs for you:

-- FIRST YOU MUST CALCULATE THE MAXIMUM NUMBER OF POSSIBLE PARTITIONS
SELECT INTEGER(MIN(4096 , POWER(2E00 , 32)
/ (
(64E00 * 1024E00 * 1024E00 * 1024E00) -- DSSIZE TS
/ (32 * 1024) -- TS PGSIZE
)
)) AS MAX_NBR_PARTS
FROM SYSIBM.SYSDUMMY1 ;
-- THEN YOU PLUG THAT NUMBER INTO HERE
SELECT MIN(
(64E00 * 1024E00 * 1024E00 * 1024E00) -- DSSIZE TS
, (POWER(2E00 , 32) / 2048 -- NBR PARTS
) * 4096 -- IX PGSIZE
) AS MAX_INDEX_SIZE
FROM SYSIBM.SYSDUMMY1 ;
-- FOR AN NPI YOU JUST NEED THE PIECESIZE AND THE INDEX PAGESIZE
SELECT INTEGER(MIN(4096 , POWER(2E00 , 32)
/ (
(8E00 * 1024E00 * 1024E00 * 1024E00) -- PIECESIZE
/ 4096 -- IX PGSIZE
)
)) AS MAX_PIECES
FROM SYSIBM.SYSDUMMY1 ;

Death by Index

Suffice it to say you can be killed by running out of space in your PI, running out of space in your NPI, or even running out of NPI LPS pieces…

A New Type of Index

A brand new index type, data-partitioned secondary index, DPSI was created in Db2 V8 so that you could access partitions without having to scan thousands of datasets. Naturally, DPSIs only work if the application SQL is cognizant of their definition and index. Failure to do so can give terrible performance. These are naturally also bound to their respective partitions. It was also at this time that NPIs got rebranded to be NPSIs…

Separation for Utilities is Good

The major utility benefit that partitioned TSs give, is the ability to run them in parallel, including RECOVER. It was a really great idea so that you would only REORG, COPY or RUNSTATS the partitions that required it, and all the partition level utilities could run in parallel.

Db2 9 was All Change

To make these REORGS faster, IBM changed the way REORG worked with partitioned objects in Db2 9. They did speed it up but it meant that you could no longer work in parallel! You could add ranges to the PARTs being reorged but still not in parallel. This caused much grief and gnashing of teeth amongst loads of customers. Thus, we developed the so-called NPI Job. Anything that has a dependency on an NPI could be squirreled away in the NPI job(s) to then be run sequentially by the job scheduler after all other jobs were finished. Nasty!

Faster, Better, Cheaper

Now, after over 35 years, the b-tree is showing its age. IBM came up with the FTB or FIT which, simply put, copies all the non-leaf pages into a private area of the bufferpool and then, using L2 cache-aware code, removes the need for I/Os, apart from the very last two, leaf and data page, and thus reduces I/O and CPU.

Limits Again

Of course this is not for *every* index. Only unique 64 bytes long or less with no IOE, TIMESTAMP with TIMEZONE, or versioning allowed. But hey, it was a start!

Duplicate Allowed

Now even duplicate indexes are there. Well, to start with, only 56 bytes or less but we are getting there! You will require PH30978 UI75643 and then set the new ZPARM FTB_NON_UNIQUE_INDEX to be YES.

Is an NPI Always an NPI?

Now to the core of the matter. As you should all be aware we can, talking about base tablespaces, only create Universal Table Spaces (UTS) in the three flavors we know and love – Partitioned by Growth, Partitioned by Range, and Partitioned by Range Relative Page Numbering.

And???

Well, imagine you have a TS with MAXPARTITIONS = 1 – by definition it cannot go to multiple partitions but all indexes created on it are also by definition NPIs. This is also sort of true for MAXPARTITIONS > 1 if the current number of defined and existing partitions (NUMPARTS) is equal to one.

“Fake” NPIs

These are therefore, in my humble opinion, fake NPIs – You could argue that the NUMPARTS = 1 is “risky”, as Db2 could add a new partition by the next INSERT/UPDATE, but I think the risk is low. For these cases I think the jobs should not go to special NPI handling.

What are your Db2 maintenance jobs doing with these fake NPIs?

I would be very interested to hear from you all about your thoughts and plans here!

TTFN

Roy Boxwell