Mark the first-screen rules once, split the stylesheet in the build, and confirm the rest loads without blocking.
Record coverage in Chrome DevTools, test the states that hide rules, and remove dead CSS in small, reversible steps.
Count the files tsc reads, narrow the include list and turn on incremental with a cache file. After this, you know which change made the build faster because you measured it.
Run one slow query with EXPLAIN ANALYZE and read it with a five-question checklist: rows touched, filtered rows, estimates, loops and sort spills. After this, you can find the slow node of a plan and check your fix.
Turn on pg_stat_statements in a Docker container and list the statements that cost the most in total. After this, you know which query to tune first, and you do not have to guess.
Build an index with CREATE INDEX CONCURRENTLY while a writer is active, and handle the transaction-block error and the invalid index that a failed build leaves. After this, you can add an index to a busy table without stopping its writers.
Index only the rows that queries ask for, compare the size with a full index, and enforce a unique rule for active rows. After this, you know why a prepared statement can ignore a partial index, and how to avoid it.
Add a GIN index to a JSONB column, filter with the contains operator, and compare it with an expression index and jsonb_path_ops. After this, you can tell which JSONB filters use an index and which scan the table.
Read pg_stat_user_indexes to find indexes that were never scanned, find identical indexes in pg_index, and drop them with DROP INDEX CONCURRENTLY. After this, you can remove indexes that cost writes and give nothing back.
Add a BuildKit cache mount to pip install so a changed requirements file downloads only the new packages. After this, a dependency change no longer means downloading everything again.
Build a Go service in a multi-stage Dockerfile and ship it in an image well under 50 MB, with a non-root user and a health check. After this, you know why Go gets that small and Node does not.
Set CPU and memory limits, watch them work, change them on a running container and set them in Compose. After this, one busy container cannot take the whole machine.