Repository navigation
Description
Summary
google-cloud-bigquery and google-cloud-bigquery-derive currently expose procedural-macro implementation details directly in google-cloud-bigquery's documented public API (google_cloud_bigquery::query) and emit calls directly to user-facing methods instead of routing generated code through a #[doc(hidden)] pub mod __private module.
Because google-cloud-bigquery was recently bumped to 1.0.0 on main (#7424, not yet released), shipping 1.0.0 as-is will freeze macro-internal types and trait methods under 1.x SemVer.
Problem Details
Per Rust proc-macro conventions (serde::__private, thiserror::__private, etc.), README.md ("any public symbol that is not part of the public API have the #[doc(hidden)] attribute set"), and doc/contributor/howto-guide-handwritten-client-api.md:
Generated code calls only a
#[doc(hidden)] pub mod __privatemodule in the runtime crate, never the user-facing API. Otherwise every method used by the generated code is frozen, andcargo-semver-checkscannot detect violations.
Currently:
- Derive-only types and methods are exposed in
google_cloud_bigquery::query:google_cloud_bigquery::query::SqlValue(src/bigquery/src/query/from_sql.rs, re-exported insrc/bigquery/src/query.rs) is not returned or accepted by any user-facingRowor client API; it only exists to pass cell data intoFromSql::from_value.SqlValue::takeexplicitly documents that it is "Used by the#[derive(FromSql)]macro to extract struct fields", yet neitherSqlValuenorSqlValue::takeis marked#[doc(hidden)].FromSqlhas a rustdoc warning ("Do not implement this trait directly. ... New methods (such as Arrow conversion methods) may be added in future versions, which would break manual implementations"), butFromSqlis not sealed via a#[doc(hidden)]supertrait andFromSql::from_value(value: SqlValue)is not#[doc(hidden)]. Prose warnings do not exclude public traits or methods from SemVer orcargo-semver-checks.
- Generated code calls user-facing methods and unrooted paths directly (
src/bigquery-derive/src/lib.rs):#[derive(FromRow)]expands torow.take(...)ongoogle_cloud_bigquery::query::Rowand spellsstd::convert::TryFrom,std::result::Result,google_cloud_bigquery::query::Row, andgoogle_cloud_bigquery::error::RowErrorwithout leading::.#[derive(FromSql)]expands tovalue.take(...)ongoogle_cloud_bigquery::query::SqlValueand spellsgoogle_cloud_bigquery::query::{FromSql, SqlValue}andgoogle_cloud_bigquery::error::ConvertErrorwithout leading::.- Using dot-call syntax (
row.take(...),value.take(...)) on unprefixedrow/valuelocals instead of fully qualified function calls in__privatefreezes user-facing method signatures and risks method-resolution ambiguity if an in-scope extension trait defines.take(...). - Generated
implblocks are not enclosed inconst _: () = { ... };.
- Crate-level rustdoc on
google-cloud-bigquery-derivepublishes maintainer policy:src/bigquery-derive/src/lib.rsuses//!doc comments for internal# Compatibility/## Contract testsCI instructions instead of//comments, and does not advise users to importFromRow/FromSqlfromgoogle_cloud_bigquery::queryrather than depending ongoogle-cloud-bigquery-derivedirectly.
- Version skew note from chore(bigquery): bump to 1.0.0 #7424:
- chore(bigquery): bump to 1.0.0 #7424 bumped
google-cloud-bigqueryfrom0.19.0to1.0.0while leavinggoogle-cloud-bigquery-deriveat0.2.0. When migratinggoogle-cloud-bigquery-deriveto call::google_cloud_bigquery::__private,google-cloud-bigquery-derivemust also take a breaking bump (0.3.0or1.0.0) sogoogle-cloud-bigquery 0.19.0(which depends on^0.2.0) does not resolve a0.2.xpatch release that references__private.
- chore(bigquery): bump to 1.0.0 #7424 bumped
Expected Behavior
- Add
#[doc(hidden)] pub mod __privateat the root ofgoogle-cloud-bigqueryexposing only the internal types, traits, re-exports (Result,TryFrom), and helper functions needed bygoogle-cloud-bigquery-derive. - Remove
SqlValuefromgoogle_cloud_bigquery::query's public exports. - Keep
google_cloud_bigquery::query::FromSqlpublicly nameable as a trait bound (forRow::get<T: FromSql>,Row::take<T: FromSql>, and generic callers), while sealing it / movingfrom_valueonto a#[doc(hidden)]supertrait in__privateso manual external implementations are prevented at the type/doc level rather than only by a prose warning. - Update
google-cloud-bigquery-deriveto emitconst _: () = { ... };blocks referencing only::core::.../::google_cloud_bigquery::__private::...via fully qualified function calls. - Convert the maintainer
# Compatibilitysection insrc/bigquery-derive/src/lib.rsfrom//!to//comments and add brief user-facing crate docs pointing togoogle_cloud_bigquery::query::{FromRow, FromSql}. - Take a breaking version bump on
google-cloud-bigquery-derivealongsidegoogle-cloud-bigquery1.0.0.
Activity
dbolduc commentedon Nov 11, 2026
I am pretty sure that (1) is bogus.
pub mod __internalat all costs, because it is still part of the public API. I have to think it is preferable to use the real public APIs.SqlValue::take()should not be hidden. Applications can and should use them. Our macro just helps them save a lot of boilerplate code.__privatetrait, but that has the same problems of the first bullet.Most of (2) seems legit. It seems like we should harden the code with the fully qualified names. And TIL the const impl block trick. This stuff feels more like a cleanup to me.
(3) is accurate, but a harmless clean up.
(4) is only applicable if we do (1) which I don't think we should do.