LearnPerformance
How To Read Wait Statistics Without Guessing
2 min readUpdated 6 Nov 2026
Wait statistics are the fastest way to turn 'the server is slow' into 'the server is waiting on disk'. Most people misread them because half of the list is harmless background noise.
Filter the harmless waits first
The wait stats view collects everything the engine waits on, including waits that mean nothing is wrong. Idle waits, broker waits, trace and event-processing waits, and memory-manager bookkeeping waits should be excluded before you rank anything.
- Sleep, broker, service broker and queue waits — the engine waiting for work.
- Trace, event-processing and pre-replication waits — background plumbing.
- Memory grant and memory-queue waits, unless they dominate the list.
Rank by percentage of total wait time
After filtering, order by wait time and look at the top five. Those five usually explain most of the day. Percentages matter more than raw milliseconds, because raw numbers grow with uptime.
- PAGEIOLATCH_SH — reads waiting for data pages: not enough memory, or too many pages being read.
- CXPACKET — parallel queries coordinating: sometimes normal, sometimes a sign of a missing index.
- LCK_M_* — sessions queued behind locks: a blocking problem, not an I/O problem.
- WRITELOG — commits waiting on the log disk: usually log placement or antivirus scanning.
- SOS_SCHEDULER_YIELD — CPU pressure: too much work per core, or a runaway query.
Waits say what, not who
A wait type names the resource, never the query that caused it. The next step is always the same: find the sessions running during the spike, then their queries, then their plans.
Compare the same window yesterday and today. A wait type that has always been first is your baseline; a wait type that jumped is your incident.
What to remember
- Filter background waits before ranking anything.
- Read the top five by percentage, not raw milliseconds.
- Every wait type points at a resource, not a query.
- Compare against yesterday before acting on today.
