An operation's text is printed compact

Status: accepted, 2026-10-06. Answers #36. Serves The compiler decides and Relay's words. Reopen if a conforming server refuses compact text, which would be the server's defect, or if the one-line text costs a reader more than it saves on the wire; the remedy is then a formatted companion where people read, never a change to what is sent.

Context#

The compiler prints each operation's text with Relay's printer and its default options: one field a line, two spaces an indentation level, a blank line before each fragment. The runtime sends that text as it is, in every request of an operation that has no persisted id. The issue measured the text at about 2.4 times what the same operation needs, and a server refused a 65 KB operation with HTTP 400 that it accepted at 33 KB with the insignificant whitespace removed. Whitespace and commas are insignificant in GraphQL, so the server's answer cannot change with them. Relay's printer has the other form already, PrinterOptions::compact, and Relay's compiler applies it to every artifact under the feature flag compact_query_text, off by default. persistConfig takes the text off the wire altogether, but only for a server with a registry; the issue's server has a limit and no registry.

Decision#

• Every operation's text is printed compact, with Relay's printer's own option: no newline, indentation or optional space, a comma between items, strings and block strings as they are. The fragments the operation reaches follow it in the same text, as before. There is no option, and nothing at run time touches the text.
• The one text is what the artifact holds, what a request sends, what a persisted id is hashed from, what the persisted file maps the id to, what batonc print prints and what spec/documents keeps. Nothing else prints a second form of it.
• The artifact's literal is one line, unless the text has a newline: Relay's compact printer keeps one between selections that share repeated @include or @skip conditions, and the literal then spans lines as it always did.
• A fragment's source definition in the report stays as the author wrote it, formatted. It is a compile input for a dependent target, whose compiler points its diagnostics into it, not a wire payload.

Evidence#

• The goldens at this commit against the commit before, a deterministic count: the test target's 115 operations hold 61,797 bytes of text where they held 83,885, 74%; the realistic documents, without the corpus of hostile names, 18,916 where they held 26,855, 70%; the Fixture query, the deepest of them, 283 where it held 607, 47%. The gain grows with depth, since indentation does, which is where a server's limit is met.
• The issue's measurements on four documents of nested unions: compact text at 40% to 43% of the printed, and the 65 KB operation accepted at 33 KB.
• Relay's printer at the pinned revision, read 2026-10-06: compact skips the newline where it may, drops optional spaces, writes , as the item separator, and prints a string constant as it is.
• Nothing is measured at run time, and nothing needs to be: no read or commit changes, and the runtime does not read the text.

Not chosen#

• Relay's flag, featureFlags.compact_query_text, off by default: a knob with no case for its off position here. Relay keeps the formatted text as its default so that persisted ids registered under it keep working, a debt Baton does not owe before 1.0, and the block would be the first feature flag in baton.json.
• Removing the whitespace in the transport, the issue's workaround: the runtime never parses GraphQL, and the id would no longer be the hash of the text sent.
• Compact on the wire and formatted in the artifact: two texts, and the id hashes one of them.
• A pass of our own after Relay's printer, for the newline it keeps in a group of conditions: a tokenizer for a few bytes, and a string argument holding a newline would be at risk.