tl;dr – I do not recommend this book.
I was so incredibly excited when it was originally announced. A book published by VMware Press, written by prominent VMware, SQL, and storage consultants? GREAT! So much has changed in those topics over the last few years, and it’s high time we got official word on how to do a great job with this combination of technology. Everybody’s doin’ it and doin’ it and doin’ it well, so let’s get the best practices on paper.
When it arrived on my doorstep, I did the same thing I do with any new tech book: I sit down with a pad of post-it notes, I hit the table of contents, and I look for a section that covers something I know really well. I jump directly to that and I fact-check. If the authors do a great job on the things I know well, then I’ve got confidence they’re telling the truth about the things I don’t know well.
I’ll jump around through pages in the same order I picked ’em while reading:
Page 309: High Availability Options
Here’s the original. Take your time looking at it first, then click on it to see the technical problems:
OK, maybe it was bad luck on the first page. Let’s keep going.
Page 111: Database File Design
The “Microsoft Recommended Settings” are based on a 2006 article about Microsoft SQL Server 2005. I pointed this out to the book’s authors, who responded that Microsoft’s page is “published guidance” that they still consider to be the best advice today about SQL Server performance. Interesting.
Even so, the #3 tip in that ancient Microsoft list is:
3. Try not to “over” optimize the design of the storage; simpler designs generally offer good performance and more flexibility.
The book is recommending the exact opposite – a minimum of one data file per core for every single database you virtualize. That’s incredibly dangerous: it means on a server with, say, 50 databases and 8 virtual CPUs, you’ll now have 400 data files to deal with, all of which will have their own empty space sitting around.
I asked the authors how this would work in servers with multiple databases, and they responded that I was “completely wrong.” They say in a virtual world, each mission critical database should have its own SQL Server instance.
That doesn’t match up with what I see in the field, but it may be completely true. (I’d be curious if any of our readers have similar experiences, getting management to spin up a new VM for each important database.)
So how are you supposed to configure all those files? Let’s turn to…
Page 124: Data File Layout on Storage
Imagine this setup for a server with dozens of databases. And imagine the work you’d have to do if you decide to add another 4 or 8 virtual processors – you’d have to add more LUNs, add files, rebalance all of the data by rebuilding your clustered indexes (possibly taking an outage in the process if you’re on SQL Server Standard Edition).
What’s the point of all this work? Let’s turn to…
Page 114: You Need Data Files for Parallelism
No, you don’t need more data files for parallelism. Paul Randal debunked that in 2007, and if anybody still believes it, make sure to read the full post including the comments. It’s simply not true.
I asked the authors about this, and they disagree with Paul Randal, Bob Dorr, Cindy Gross, and the other Microsoft employees who went on the record about what’s happening in the source code. The authors wrote:
You can’t say Microsoft debunked something when they still have published guidance about it…. If in fact if your assertions were accurate and as severe then we would have not had the success we’ve had in customer environments or the positive feedback we’ve had from Microsoft. I would suggest you research virtualization environments and how they are different before publishing your review.
(Ah, he’s got a point – I should probably start learning about SQL on VMware. I’ll start with this this guy’s 2009 blog posts – you go ahead and keep reading while I get my learn on. This could take me a while to read all these, plus get through his 6-hour video course on it.)
So why are the authors so focused on micromanaging IO throughput with dozens of files per database? Why do they see so many problems with storage reads? I mean, sure, I hear a lot of complaints about slow storage, but there’s an easy way to fix that. Let’s turn to page 19 for the answer:
Page 19: How to Size Your Virtual Machines
Ah, I think I see the problem.
To make matters worse, they don’t even mention how licensing affects this. If you’re licensing SQL Server Standard Edition at the VM guest level, the smallest VM you can pay for is 4 vCPUs. Oops. You’ll be paying for vCPUs you’re not even using. (And if you’re licensing Enterprise at the host level, you pay for all cores, which means you’re stacking dozens of these tiny database servers on each host, and managing your storage throughput will be a nightmare.)
In fact, licensing doesn’t even merit a mention in the Table of Contents or the book’s index – ironic, given that it’s the very first thing you should consider during a virtual SQL Server implementation.
In Conclusion: Wait for the Second Edition
I’m going to stop here because you get the point. I gave up on the book after about fifty pages of chartjunk, outdated suggestions, and questionable metrics (proc cache hit ratio should be >95% for “busy” databases, and >70% for “slow” databases).
This is disappointing because the book is packed with information, and I bet a lot of it is really good.
But the parts I know well are not accurate, so I can’t trust the rest.