Crime Scene Investigation: SQL Server

“The packages are running slower in Prod than they are in Dev”

My week began with this simple declaration from one of our lead BI developers, quickly followed by an emailed spreadsheet demonstrating that, over 5 executions, an extensive ETL process was running average 630 seconds faster on Dev than on Prod. The situation needed some scientific investigation to determine why the same code, the same data, the same schema would yield consistently slower results on a more powerful server. Prod had yet to be officially christened with a “Go Live” date so I had the time, and having recently been binge watching CSI: New York, I also had the inclination.

An inspection of the two systems, Prod and Dev, revealed the first surprise: although Prod was indeed a “bigger” system, with double the amount of RAM of Dev, the latter actually had twice as many processor cores. On neither system did I see much sign of resources being heavily taxed, while the ETL process was running. Without any real supporting evidence, I jumped to a conclusion that my years of performance tuning should have helped me avoid, and that was that the hardware differences explained the better performance on Dev.

We spent time setting up a Test system, similarly scoped to Prod except with 4 times the cores, and ported everything across. The results of our careful benchmarks left us truly bemused; the ETL process on the new server was slower than on both other systems. We burned more time tweaking server configurations, monitoring IO and network latency, several times believing we’d uncovered the smoking gun, until the results of subsequent test runs pitched us back into confusion.

Finally, I decided, enough was enough. Hadn’t I learned very early in my DBA career that almost all bottlenecks were caused by code and database design, not hardware? It was time to get back to basics. With over 100 SSIS packages and hundreds of queries, each handling specific tasks such as file loads, bulk inserts, transforms, logging, and so on, the task seemed formidable. And yet, after barely an hour spent with Profiler, Extended Events, and wait statistics DMVs, I had a lead in the shape of a query that joined three tables, containing millions of rows, returned 3279 results, but performed 239K logical reads. As soon as I looked at the execution plans for the query in Dev and Test I saw the culprit, an implicit conversion warning on a join predicate field that was numeric in one table and a varchar(50) in another! I turned this information over to the BI developers who quickly resolved the data type mismatches and found and fixed “several” others as well. After the schema changes the same query with the same databases ran in under 1 second on all systems and reduced the logical reads down to fewer than 300.

The analysis also revealed that on Dev, the ETL task was pulling data across a LAN, whereas Prod and Test were connected across slower WAN, in large part explaining why the same process ran slower on the latter two systems. Loading the data locally on Prod delivered a further 20% gain in performance.

As we progress through our DBA careers we learn valuable lessons. Sometimes, with a project deadline looming and pressure mounting, we choose to forget them. I was close to giving into the temptation to throw more hardware at the problem. I’m pleased at least that I resisted, though I still kick myself for not looking at the code on day one. It can seem a daunting prospect to return to the fundamentals of the code so close to roll out, but with the right tools, and surprisingly little time, you can collect the evidence that reveals the true problem. It is a lesson I trust I will remember for my next 20 years as a DBA, if I’m ever again tempted to bypass the evidence.