There is an interesting case in this week. A query run over hours, it take millions of buffer gets. The query is to verify if there is any data in the past 3 month for a customer, which usually completes in a minute. The case can’t be reproduced in production clone DB. The same execution path return around 10 seconds, doing 3829 buffer gets. The execution plan is as expected as below, the index skip scan will return hundreds of rows, zero or several rows return after nested loop join. Continue reading →