Memory-optimized TempDB tables are the system tables that SQL Server 2019 and later can keep in memory. This post shows two queries: one that checks whether the feature is on, and one that lists the tables it covers. It also explains the limits that come with the feature.
Metadata Contention in TempDB: When Memory-Optimized Metadata Helps
Metadata contention in TempDB happens when many sessions create and drop temporary tables at the same moment. This post reads the waits that point to it, explains why RESOURCE_SEMAPHORE is a different problem, and shows how SQL Server 2019 and later can keep that metadata in memory.
Visible Offline Schedulers: Find CPUs SQL Server Ignores
Visible offline schedulers are CPUs that Windows shows and SQL Server refuses to use. This post reads the scheduler status with tested queries, compares it with the CPU count and affinity, and explains the two common causes. It also shows the statement that returns the CPUs to automatic affinity.
Nonclustered Primary Key Test: Last Page Insert Waits
A nonclustered primary key test shows whether a heap fixes last page insert waits. On SQL Server 2025 it doesn’t, because the key index still rises. A lane layout that spreads the inserts removes the latch wait, and the post shows the cost of that layout.
OPTIMIZE_FOR_SEQUENTIAL_KEY: Test the Last Page Insert Fix
OPTIMIZE_FOR_SEQUENTIAL_KEY is an index option that targets last page insert contention. This post tests it with 32 connections that insert into a clustered identity key, with the option off and on. The option changes the kind of waiting, and the post shows how to decide whether it helps your table.
SQL SERVER – Last Page Insert PAGELATCH_EX Contention Due to Identity Column
Today’s goal is to understand what is the Last Page Insert PAGELATCH_EX Contention Due to Primary Key. Let us start understanding it together.
Recovery Model Insert Test: FULL vs SIMPLE Under Load
A recovery model insert test compares FULL and SIMPLE under a load of 16 connections. The log written is almost the same in both, and the elapsed time depends more on the server than on the model. The real difference is when the log can be reused, and this post shows that step by step.







