SQL Server string statistics

Tue, 20 Oct 2026 00:00:00 +0000

Co-written with Claude, who ran much of the testing and research.

TL;DR

String statistics have existed in SQL Server since version 2005. They are used to estimate the cardinality of LIKE searches, such as Email LIKE '%.org'. When the first column of a statistics object is a character type, that is char, varchar, nchar or nvarchar, including varchar(max), nvarchar(max), text and ntext, the statistics object holds a string summary in addition to the histogram. You can see it in DBCC SHOW_STATISTICS, where the String Index property is then YES.

This summary is a sample of the column values, taken at a regular interval in sort order. Its size is 5% of the rows read by the statistics, plus 10; with FULLSCAN, 5% of the table rows: 510 strings for 10,000 rows, 2,510 for 50,000. But from about 56,000 rows read on, it stops growing and stays around 2,800 strings, whatever the size of the table (and even fewer for nvarchar and for long strings). The strings are stored in a trie (prefix tree): each node holds a piece of a string, and strings that start the same way share the nodes of that beginning.

To estimate a LIKE with wildcards, the optimizer applies the pattern to each string of the sample, then multiplies the proportion found by the number of rows in the table. When no string in the sample matches the pattern, the proportion would be zero, and so would the estimate. The optimizer then acts as if a tenth of a string matched: on the million rows of Contact.ProspectUS, whose sample holds 2,825 strings, a pattern not found in the sample is estimated at 1,000,000 × 0.1 / 2,825, or 35.4 rows. That is the floor of the estimate.

Frequent patterns are estimated correctly, and rare patterns get the floor value, roughly the number of rows divided by 28,000. The summary is not used for cardinality estimation when the column is wrapped in an expression, such as a function call, when the pattern uses ESCAPE, comes from a local variable or is nvarchar against a varchar column (hence an implicit conversion), or when the column is a varchar with a UTF-8 collation. In those cases, the current estimator guesses 9% of the table rows for a pattern that starts with a wildcard, and estimates a prefix from the histogram. Beyond 80 characters, only the first 40 and the last 40 are kept, joined together.

Statistics on Rudi Bruchez

SQL Server string statistics