SuperAPI AI Directory

pg-code-style-guide

General & other

View repository

Updated 10-16 07:19

About this skill

src/postgres code style guide

Everything under src/postgres is a fork of upstream PostgreSQL that gets re-merged on every PG major-version upgrade. The PG 11->15 merge hit 600+ conflicts (~2.3 per work day for a year), and every conflict avoided is catchup time saved. Every rule here serves one of two goals:

  1. Keep the YB delta small, clustered, and visibly marked so a merger (human or git) can tell YB code from PG-owned code at a glance.
  2. Keep PG-owned lines byte-identical to upstream so git auto-merges them and conflicts appear only where a real logical conflict exists.

The pinned upstream version of any file is defined by src/lint/upstream_repositories.csv. To see the YB delta of a file, or to fetch the pinned upstream copy itself:

src/lint/diff_file_with_upstream.py <file> -b
src/lint/upstream_file.py <file> <output-path>

Older files predate these conventions and violate them freely (e.g. createplan.c is full of interleaved YB code). Surrounding code is NOT precedent: follow the guide for the lines you add or change. Do not go fixing violations elsewhere in the file -- that is a refactor and belongs in a cleanup-only revision (below) -- but when your change lands right on one (say, appending declarations to a YB cluster that never got its /* YB declarations */ opener), fixing it in place is allowed.

Add a yb prefix

Put yb in the name of every YB-introduced identifier that is visible from upstream context: functions, global and file-scope variables, struct fields, types, macros, GUC variables, parameters added to upstream functions. Normally yb is the prefix; where a fixed prefix is imposed, it comes right after (pg_yb_... catalog names, _YB_... keyword tokens). Match the casing style of the surroundings. The marker is what lets a merger (and lint) attribute a line to YB without archaeology.

  • Markers flag the YB/PG boundary, not every line inside it. Within code that is already YB by context -- the body of a yb-prefixed function (yb_foo, is_yb_foo, YbFoo, ...), a YB-introduced file, or a block gated by a YB condition (IsYugaByteEnabled(), IsYBRelation(...), a yb_* GUC) -- local variables take plain PG-style names and comments take no YB: prefix: the enclosing context already attributes them, and marking some lines but not others is noisier than marking none. Use /* YB: ... */ markers only where YB code sits directly in PG-owned context and the marker is what attributes it.
  • Do not introduce new ybc-prefixed names: that namespace belongs to pggate, whose YbcPg* types src/postgres code uses everywhere. (ybctid is not one of them: it parses as yb + ctid.) When extending an existing ybc-named family (e.g. the ybcin* index-AM handlers), match the family rather than stick out.
  • YB-introduced files must be recognizable by name: start the file name with yb (pg_yb only where a constraint forces it, such as catalog table filenames), and every type they define must contain yb (any casing: yb/Yb/YB). File names take the casing of the upstream family they join -- executor nodes are camel, nodeHashjoin.c -> nodeYbBatchedNestloop.c, today the only such family (lint allows Yb in a file name there and nowhere else). YB-introduced files must NOT contain a /* YB includes */ block -- the whole file is YB, so includes are just normal includes.

Cluster YB code

Group YB code separately from PG code. Upstream constantly appends to its own lists -- includes, declarations, struct fields, function parameters, functions -- and a YB line sitting inside one turns every upstream append into a conflict. Do not let a YB addition sit directly against a PG-owned line: put it in a YB block opened by a marker comment and separated from PG code by a blank line.

  • /* YB includes */ -- one block after the upstream includes, holding every YB-added include; lint enforces the exact format and the postgres.h/c.h goes-first exception.
  • /* YB declarations */ -- for a block of YB static/extern declarations. The block goes at the end of the upstream declaration list -- in a header, at the bottom of the file (before the closing #endif); append to the YB block already there if one exists. Declarations stay bare: PG puts the descriptive comment on the definition, not the declaration.
  • /* YB fields */ -- for YB fields appended to a PG-owned struct, enum, or union (append at the end unless field order is semantically constrained). Keep PG's member-comment split: at most a short one-liner on the member itself; a longer description goes in a YB: paragraph appended at the bottom of the type's header comment, before the closing */ -- the upstream comment lines themselves stay untouched.
  • Functions: place new YB functions at the bottom of the file (or in an existing YB function cluster), not between two upstream functions. If a static YB function is called from code above it, add a prototype to the /* YB declarations */ block instead of moving the definition up. Exception: a yb function factored out of an upstream function stays where the upstream code was when that keeps the diff vs upstream smaller.
  • Function parameters: cluster YB parameters at the end of the upstream parameter list and yb-prefix them; do not insert into the middle if avoidable.

Interleaving is otherwise allowed only when an ordering constraint forces it -- a sorted list (e.g. kwlist.h keywords) or semantically ordered entries. Each interleaved YB entry then carries its own marker: a yb-prefixed identifier (as in _YB_ACCOUNT_P) or, if the entry has no identifier, an inline /* YB */ comment.

Convention clusters YB blocks at the bottom, after the corresponding upstream section. Upstream tends to append its own new code there too, so conflicts at the seam are expected. The marker comment and the blank line are what make them quick to attribute and resolve -- never skip them.

Do not touch PG-owned lines

Try to keep PG-owned lines exactly byte-identical. The one allowance, when a YB change must wrap them, is an indentation-only change -- diff_file_with_upstream.py -b and merge tooling can see through that.

  • Do not fix upstream style, typos, or whitespace -- if the style is wrong, leave it wrong. Trailing-whitespace "cleanup" of PG-owned lines has caused real merge conflicts.
  • When a YB if/else must wrap PG-owned lines, do a pure indent: add leading tabs only, and do not rewrap the lines to fit 80 columns -- rewrapped lines defeat the indentation-blind tooling.
  • Avoid refactors and style changes of PG code in a feature diff, especially while a PG major-version merge is ongoing (save them for after it completes). A needed code move gets its own revision, as a pure move separate from any modification, so git rename detection and bisect keep working.

Avoid brittle tests

The regress test file taxonomy -- upstream-named files, yb.port.*, yb.orig.*, yb.depd.*, schedule ordering, YB-deviation marking -- is documented in src/postgres/src/test/regress/README. Read it before touching test files. Do not edit a test with an upstream name to make a YB behavior pass: the YB difference belongs in the yb.port.* version, marked with a -- YB: <reason> comment. (Lint pins upstream-named files to src/lint/upstream_repositories.csv, which also scopes exceptions such as pg_hint_plan.)

Avoid test assertions that you expect will break on a PG major-version upgrade:

  • Never assert absolute catalog versions.
  • OIDs: [8000,10000) are YB-owned and never change; [10000,16384) are dynamically allocated by initdb and change; >= 16384 are user-allocated and change for many reasons. OIDs < 8000 can in some cases change, but you can assume they won't and hardcode them if you must.
  • Prefer structural or relative assertions (row counts, plan shapes, ratios) over environment-dependent absolutes.

Better documentation

If the merger knows your intention, conflict resolution gets easier. Say WHY the YB change exists, and put the why where it belongs:

  • In a code comment when it holds into the future: an invariant, the reason an expected output differs from upstream.
  • In the commit message when it holds only at that moment: the motivation, what was verified.
  • Make test plans easy for a stranger to run: specific commands, exact test names (no typos or partial specifications), no jargon, no assumed local setup.

Cleanup-only revisions

Bringing an existing file up to these conventions -- adding missing yb prefixes, YB: comments, and marker blocks, moving YB functions to the bottom -- is legitimate work, but it is a refactor: make it a dedicated revision, never part of a feature diff. Cleanup has its own rules:

  • The scope is the YB delta only. PG-owned lines stay byte-identical: upstream style, typos, and whitespace remain out of bounds even in a cleanup pass.
  • Stay behavior-preserving and mechanically checkable. Moved code must be byte-identical at its new location so diff tooling can prove the move pure; a rename updates the definition and its references and nothing else. If code needs both moving and editing, split them into separate revisions.
  • Verify that diff_file_with_upstream.py output changed only in the cleaned aspects, that the lint warning count went down, and that the file still builds and passes its regress tests.
  • During an active PG major-version merge, coordinate with the merge owners before cleaning up: cleanup churns exactly the lines the merge is resolving, so one-sided style fixes manufacture conflicts.