Build And Run

build-and-run

Slash-command wrappers (use these first)

The recipes in this skill are also exposed as slash commands under .claude/commands/. Prefer them — they encode the right flags and guardrails:

• /setup-pg — first-time meson setup + build + install (idempotent; --force to reconfigure).
• /pg-start, /pg-stop, /pg-restart — lifecycle.
• /pg-psql — psql -h /tmp -d postgres against the dev cluster.
• /pg-test [--suite NAME] [--test PAT] — meson test, always prepends --suite setup so the initdb-template trap can't bite.
• /pg-attach — get a backend PID and print the lldb -p line.
• /pg-tail-log — follow dev/data-debug/server.log.
• /pg-fresh --yes — wipe data-debug/, re-initdb (preserves the build).
• /pg-reclone-dev — nuclear: re-clone the whole dev tree from the read-only reference.
• /setup-pg-asan / /pg-start-asan — sibling profile built with AddressSanitizer + UndefinedBehaviorSanitizer (see "Sanitizer builds" below); use this when chasing memory bugs or undefined behavior.

The rest of this file is the underlying mechanics, for cases where you need to deviate from the wrappers.

Running these from a git worktree (gotcha)

Every /pg-* command above resolves paths via $PWD/dev/... and $PWD/source/.... The dev and source symlinks live at the root of the postgres-claude repo, and git worktree add (or this project's EnterWorktree tool) does NOT propagate them into the new worktree. Inside a worktree, the commands break with No such file or directory.

Two ways to handle it:

1. Run from main (simplest). The pg-* commands are stateful — they touch the dev cluster, not the meta-repo source — so running them from main is harmless even when your edits live on a feature branch.
2. Add symlinks per worktree (when you genuinely need everything in one tree). One-liner from inside the worktree, using absolute paths so they don't drift:

   ln -s /Users/matej/Work/postgres/postgresql-dev dev
   ln -s /Users/matej/Work/postgres/postgresql     source
   

   These are local-only — they won't be tracked by git (the parent .gitignore excludes them) and ExitWorktree won't try to commit them.

Sanitizer builds (build-asan profile)

The default dev/build-debug/ enables Asserts + debug symbols but no sanitizers. For memory-bug hunting (use-after-free, heap-buffer-overflow, double-free, integer/signed overflow, alignment violations) build a parallel dev/build-asan/ profile with ASan + UBSan turned on:

meson setup dev/build-asan dev \
  --buildtype=debug \
  -Dcassert=true \
  -Ddebug=true \
  -Db_sanitize=address,undefined \
  -Db_lundef=false \
  -Dprefix=$PWD/dev/install-asan
ninja -C dev/build-asan
ninja -C dev/build-asan install

-Db_lundef=false is required on macOS / clang because ASan needs runtime symbols resolved late; without it linking fails with "undefined symbols for arch arm64: ___asan_init…".

To run the cluster against this build, point PATH/PGDATA at the asan install + a separate data dir (so you can hop between debug + asan clusters without initdb collisions):

export PATH="$PWD/dev/install-asan/bin:$PATH"
export PGDATA="$PWD/dev/data-asan"
[ ! -f "$PGDATA/PG_VERSION" ] && initdb -D "$PGDATA" --locale=C --encoding=UTF8
# ASan-specific runtime knobs:
export ASAN_OPTIONS="abort_on_error=1:detect_leaks=0:detect_stack_use_after_return=1:print_stacktrace=1"
export UBSAN_OPTIONS="print_stacktrace=1:halt_on_error=1"
pg_ctl -D "$PGDATA" -l "$PGDATA/server.log" start

/setup-pg-asan and /pg-start-asan wrap this exactly.

macOS-specific caveats

• LeakSanitizer is not available on Darwin. detect_leaks=0 above is required; setting detect_leaks=1 errors out at runtime. For real leak detection you have three options:
  1. Build the asan profile on Linux (container or VM) — LSan works there.
  2. Use macOS-native leaks <pid> against the running backend (no rebuild needed). Set MallocStackLogging=1 before pg_ctl start for proper backtraces; see the debugging skill.
  3. Use pg_backend_memory_contexts to watch a single context grow across a workload (also in the debugging skill).
• ASan still catches use-after-free, heap-buffer-overflow, and double-free on macOS, which is the common case for backend C bugs.
• The ASan slowdown is ~2-3x on PG workloads — fine for development, obviously not for benchmarking.

Picking which profile to use

Repo split — important

There are TWO upstream clones in the parent dir, both symlinked into this meta repo:

• source/ → ../postgresql/ — read-only reference. Kept exactly in sync with upstream master. Used for code citations in knowledge/... docs. Never build or run from here.
• dev/ → ../postgresql-dev/ — mutable test field. All build artifacts (build-debug/, install-debug/, data-debug/) live inside here. Safe to delete and re-clone (rm -rf postgresql-dev && git clone --no-hardlinks postgresql postgresql-dev). This is also where any local patches / experimental branches live.

When citing source code in docs use source/... paths (stable). When running commands that build, install, initdb, or modify the tree, use dev/....

Toolchain choice — meson is the default

Since PG 16, meson is the primary build system. autoconf still works and the buildfarm still exercises it, but new instructions assume meson unless told otherwise.

One-time setup (debug build)

From postgres-claude/:

# out-of-tree build dir, kept beside dev/'s source
meson setup dev/build-debug dev \
  --buildtype=debug \
  -Dcassert=true \
  -Ddebug=true \
  -Dprefix=$PWD/dev/install-debug

ninja -C dev/build-debug
ninja -C dev/build-debug install

Notes:

• cassert=true turns on the Assert() macros — invaluable when reading internals.
• --buildtype=debug (not debugoptimized) keeps symbols readable in gdb/lldb.
• Build dir is dev/build-debug so it sits inside the mutable clone and is ignored upstream by build*/. The dev clone's .git/info/exclude also ignores build-debug/, install-debug/, and data-debug/ so PG-relative status is always clean.

macOS — match the sibling postgresql-dev/build-debug config (F7)

On macOS, a fresh meson setup build-debug may need explicit ICU include flags to match the existing sibling configuration that all the other build trees use:

meson setup dev/build-debug dev \
  --buildtype=debug \
  -Dcassert=true \
  -Ddebug=true \
  -Dc_args='-I/opt/homebrew/opt/icu4c/include' \
  -Dprefix=$PWD/dev/install-debug

The reference configuration to mirror lives at /Users/matej/Work/postgres/postgresql-dev/build-debug — when in doubt, copy its meson-info/intro-buildoptions.json settings. Without the ICU include path, configure trips when probing for unicode/ucol.h on homebrew-arm64 setups.

Origin: sesvars calibration F7.

First-run gotcha: the setup suite must run ONCE before regress (F7)

On a freshly-created dev/build-debug/ tree (right after meson setup + ninja install), the test harness has NOT yet built tmp_install/initdb-template/. That artifact is produced by meson's setup suite (targets tmp_install, install_test_files, and critically initdb_cache). Until those run once, any --suite regress invocation fails with:

copying of initdb template failed
... tmp_install/initdb-template: No such file or directory
... could not bind IPv4 address ...: Address already in use

Run the setup suite once before the first regress run:

meson test -C dev/build-debug --suite setup
# then, normally:
meson test -C dev/build-debug --suite setup --suite regress --num-processes 4

The /pg-test wrapper prepends --suite setup for exactly this reason, but a manual meson test --suite regress outside the wrapper will silently fail without it.

Origin: sesvars calibration F7.

Initialize a cluster and start it

export PGDATA=$PWD/dev/data-debug
export PATH=$PWD/dev/install-debug/bin:$PATH

initdb -D "$PGDATA" --locale=C --encoding=UTF8
pg_ctl -D "$PGDATA" -l "$PGDATA/server.log" start
psql -h /tmp -d postgres
# ...
pg_ctl -D "$PGDATA" stop -m fast

The edit -> rebuild -> retest cycle

After editing any backend C file:

ninja -C dev/build-debug install       # incremental — ninja recompiles only what changed
pg_ctl -D "$PGDATA" restart            # CRITICAL — see below
psql -h /tmp -d postgres -c 'SELECT 1;'

Why the restart is mandatory: the running postmaster is still mapped to the previous postgres image, and every backend it forks for new connections inherits that image. ninja install only updates the on-disk binary — without a restart, your edits are invisible. The same applies to shared libraries / extensions under dev/install-debug/lib/ and dev/install-debug/share/.

Slash-command wrappers that automate this: /setup-pg (build+install), /pg-restart (stop -m fast + start), /pg-psql (psql -h /tmp). See .claude/commands/.

The per-connection fork model — critical gotcha

PostgreSQL uses a process-per-connection model. The postmaster forks a new backend process for every incoming psql connection. Implications when debugging:

• Attaching gdb to the postmaster does not let you step through query execution — query execution happens in the forked child.
• The pattern is: connect with psql, get the backend PID with SELECT pg_backend_pid();, then gdb -p <pid> in another terminal.
• Or use --single (single-user mode, no postmaster, no fork) for stepping through startup-time code paths.

Single-user mode

postgres --single -D "$PGDATA" postgres

Runs the backend in the foreground bound to your terminal. No SQL frontend, no fork. Useful for debugging code that runs early in backend startup or for running short SQL non-interactively under a debugger.

Launching single-user mode under lldb

For startup paths (InitPostgres, recovery, GUC bootstrap, shmem init) that run before any client connects, launch the backend under the debugger so you're at instruction 0 with no fork race:

lldb -- $PWD/dev/install-debug/bin/postgres --single -D "$PGDATA" postgres
(lldb) breakpoint set --name InitPostgres
(lldb) run

This sidesteps the per-connection fork problem entirely — there is no postmaster, no fork, just one process you control end-to-end.

For post-startup query debugging, attach to a live backend by PID instead — see .claude/commands/pg-attach.md, which automates the SELECT pg_backend_pid() + lldb -p <pid> flow.

Useful debug knobs

• log_min_messages = debug5 in postgresql.conf for maximum verbosity.
• log_statement = 'all' to see every statement.
• client_min_messages = debug1 per-session in psql for inline debug output.
• debug_print_parse / debug_print_rewritten / debug_print_plan for planner work.

Running the regression tests

ninja -C dev/build-debug test         # full meson test target
# Or a single suite — MUST include --suite setup so tmp_install /
# initdb-template get built first:
meson test -C dev/build-debug --suite setup --suite regress --num-processes 4

The regress suite is the headline one (~33s wall on Apple Silicon, plus ~7s for the setup suite that builds tmp_install/ and the initdb-template). isolation covers concurrency (serializable, predicate locks, etc.). recovery covers crash & replication.

Gotcha: meson test --suite regress alone fails with copying of initdb template failed because the setup suite (tmp_install, install_test_files, initdb_cache) hasn't run. Always pass --suite setup alongside any filtered suite selection.

Common build pitfalls

• Stale build dir after pulling upstream — ninja -C dev/build-debug clean and re-setup if meson.build files changed significantly.
• Linker errors after switching branches with extension SQL changes — re-install to refresh dev/install-debug/share/.
• initdb failing with locale error on macOS — use --locale=C (above) rather than the system default.

Where the build artifacts you'll touch live

• Binaries: dev/install-debug/bin/{postgres,psql,pg_ctl,initdb,…}
• Shared libraries / extensions: dev/install-debug/lib/
• Default extension SQL: dev/install-debug/share/extension/
• Per-cluster data: dev/data-debug/ (created by initdb)
• Server log: dev/data-debug/server.log

Cross-references

• .claude/skills/debugging/SKILL.md — attach lldb/gdb to a backend after this skill gets the cluster up.
• .claude/skills/psql/SKILL.md — connect to the cluster this skill started.
• .claude/skills/testing/SKILL.md — run pg_regress / isolation / TAP against this build.
• .claude/skills/error-handling/SKILL.md — interpret Assert() triggers and ereport() output produced by a cassert build.
• .claude/commands/setup-pg.md, .claude/commands/pg-start.md, .claude/commands/pg-restart.md, .claude/commands/pg-fresh.md — slash-command wrappers built on top of this skill.
• Top-level CLAUDE.md — source/ vs dev/ symlink layout.