T-SQL Tuesday #99: Counting Rows the Less-Hard-Way

We can get our row count, and min & max date values, without ever touching the actual source table!


This month’s invite courtesy of Aaron Bertrand (B | T), whose “bad habits” blog series still inspires many an impassioned debate or engaging argument discussion on a regular basis among DBAs & Developers alike.

And yes, I’m taking the easier of the two “dealer’s choice” choices — the SQL focused one.  (I’m not big on sharing/blogging personal stuff, at least not here; I may one day start another blog for that, or perhaps just occasionally post more #off-topic stuff , but for now you’ll have to be content with my stories of vehicle troubles and the occasional movie-geekery).

So, without further ado…

By the way, what is ‘ado’ and why should there be no further of it?

Accidentally apropos on many levels…

Counting Rows in Really Big Tables

Previously touched on here, tables of unusual size (TOUSes) can be tricky.  You don’t want to lock them up for a long period of time, but you often need to gather information about them (such as row count, size, range of values) to perform some kind of operational maintenance with/on them.  In particular, Aaron’s post on “counting rows the hard way” inspired me to look into this a bit more and try to come up with a clever-ish way of finding out some basic “shape of data” info without actually querying (scanning) the table itself.

To start with, it’s actually really simple to get the total row-count from a few system catalog views — Aaron’s already shown you that, so I won’t repeat.  My interest is more in questions like “How many rows match a where-clause?” or “What are the min & max values for thatColumn?”

For this post, I’ll be focusing on a particular kind of table — the “history” or “transaction” table.  The idea here is that you have a record of “every time some event happens in/to some entity”.  A very common example is audit-trail tables, which I’ve been dealing a lot with lately.  Another common example is a “transaction history” table, such as, in our new favorite MSSQL demo database WideWorldImporters, the table Warehouse.StockItemTransaction​.  It’s the 2nd largest table in the db at 260-some-thousand rows.  {The largest is a multi-million-row beast that is actually the system-versioned aka temporal table behind a “normal” table; I might build a phase-2 example around this, but not today.}  So, while our queries won’t be super slow, we’ll get enough of an idea of what’s bad & good from measuring our IO stats (with SET STATISTICS IO ON).

TL;DR: The demo script is available here; the headers below correspond to the comment-lines of the same name, but I’ve left enough commentary in the SQL itself to keep the average reader on-track, so feel free to check it out ahead of time.  But do keep reading at some point!  :o)

A. Gathering Some Intel

First up, we have good ol’ sys.sp_spaceused.  This gives you some sizing info about the table, including its row count.  There’s a “disclaimer” circulating around out there that it’s not “up to the millisecond accurate” , i.e. it might not have the most current row count if someone else is in the middle of an insert operation or whatnot.  But for all intents & purposes, you can consider it truth.

Then you have the “hard ways” that people typically use — and that Aaron, again, covered just fine in his post on the subject, so I won’t spend any more time there.

But what if I want to count rows matching a where clause (a predicate)?  And in dealing with a typical history/transaction table, the predicate is almost always “between such and such dates”.  (Of course we won’t literally use the between operator, as we have been chastised severely; we know best to use >= and < !)  Also, I want to know the MIN and MAX of said dates in the table.  Lord knows we’re gonna be doing some table-scanning.

B. Ok, Let’s Try an Index

In their benevolent wisdom, the SQL deities decided not to give us an index on WideWorldImporters.Warehouse.StockItemTransactions.TransactionOccurredWhen.

BTW, how’s that for a verbose column name?  What, TransactionDate not good enough?  I suppose it isdatetime2 after all, but still…

So we create an index on it, to see if that helps our poor “count where dates” query.  And behold, it does!  We’ve cut our # of logical reads down by about 90% (from 1900 to 200, if you’re following along in the script).  That’s fantastic, but… we can do better.  Because if the table is, say, 500 million rows instead of 260k, that’s about 400,000 logical reads, which.. could definitely suck.

C. The Better Way

Again, the script has an ode to Aaron’s query on sys.partitions/tables to get the row-count from the meta-data.  Then the real fun begins.

There’s a system DMV (or probably ‘DMF‘ – dynamic management function) called sys.dm_db_stats_histogram, which takes the table’s object_id and the index’s index_id as arguments.  It gives you, obviously enough, the statistics histogram of the statistics object corresponding to that index.  We want to store its output in a temp-table (or even a real table — go nuts!) so we can query it some more.

--For example, if our new index is index_id 7:
sys.dm_db_stats_histogram(OBJECT_ID('Warehouse.StockItemTransactions'), 7)

So we create our #StatsHist table (“hist” being an abbreviation for “histogram”, not “history”, though in retrospect that’s probably not worth the possible confusion), and we populate it with the meta-data from Warehouse.StockItemTransactions and its new index that we just created (on TransactionOccurredWhen).  Poof!  We have an easy way of showing min/max values in that column!  Well… almost.  We have to convert the variant datatype to an understandable & aggregate-able (probably a made-up word.. aggregable? aggregatable?) type.  So we add a column range_hk_proper of type datetime2 and populate it with the converted values of range_high_key from the stats-output.

There!  Now we’re cookin’ with gas.  Our min/max/count query, and our “count where date-range” query, run in mere milliseconds, without ever touching the actual source table.  So we don’t lock it up or block anybody else from writing to it, even in the most pessimistic isolation levels.

Except when you created that index we needed on the date column.

Yes, I know.  What we’re hoping is that the tables we deal with in the “real world” already have such an index that we can take advantage of.  If not, well, that’s what maintenance windows are for.  And you better believe you’re gonna need that index sooner or later.

Been a while since I used a Family Guy meme….

Where To Next?

Ostensibly, this whole thing could probably be turned into a stored-proc so you could run it “on demand” for any table that had a date or datetime column which you wanted to get such information about.  It’d have to do a lot of error-checking, of course — it wouldn’t work if you don’t have such a column, and if there’s no index on it, and probably a myriad of other ‘gotchas’ that I’m not thinking of at the moment.  But I did try to lay the groundwork for improvement. #StatsHist stores schema & table name too, so if you felt like turning it into a mini-data-warehouse holding a BUNCH of stat-histograms for a whole mess of tables, you could definitely do that.  And then you could run some basic analytics on it — min/max/avg, counts by year/month/day, etc.

Sounds like fun, no?  ;o)

Author: natethedba

I'm a SQL Server DBA, family man, and all-around computer geek.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s