@pothos/plugin-drizzle 0.18.1 → 0.19.0

package/CHANGELOG.md

@@ -1,5 +1,350 @@
 # @pothos/plugin-drizzle
 
+## 0.19.0
+
+### Minor Changes
+
+- 0c7565a: Selections can be built by async callbacks. New `AsyncSelections` flag in the schema types, off by
+  default.
+  - `AsyncSelections: true` widens what the selection callbacks accept: a field's `select` function,
+    a relation `query` callback, a `relationCount` / `relatedCount` `where` callback, and the
+    `select` / `query` callbacks of `prismaConnectionHelpers` and `drizzleConnectionHelpers` may all
+    return promises. Without the flag they are typed as synchronous and an async callback is a type
+    error, which is what the prisma plugin needed: it accepted an async `select` and silently
+    dropped the selection it resolved to.
+  - The plugins still build one query. Callbacks start in the same tick, and what they return is
+    merged after every synchronous selection, in document order. Async argument mappers (such as the
+    validation plugin's) are awaited before the field's `select` runs. Synchronous selection planning
+    remains synchronous and does not create promises solely to support async callbacks.
+  - Inside an async `select`, `await` the result of `nestedSelection` before adding it to the
+    selection. A `select` that returns while a nested selection it started is still pending throws,
+    naming the field.
+  - Everything the plugins generate settles its plan before your resolver runs, so a schema built
+    only from `t.relation`, `t.relationCount` / `t.relatedCount`, `t.prismaField` /
+    `t.drizzleField`, `t.prismaConnection` / `t.drizzleConnection` and `t.relatedConnection` can
+    turn the flag on and change nothing else. The `query()` builder handed to a `drizzleField` or
+    `drizzleConnection` resolver never returns a promise for the same reason.
+  - The places where your own code asks for a query or a selection are the ones to audit:
+    `queryFromInfo` (prisma), a connection helper's `getQuery` (both plugins), `nestedSelection`,
+    and `mergeNestedSelection`. `queryFromInfo` and `getQuery` build the query synchronously and
+    take an `awaitSelections` option; pass `awaitSelections: true` to get a `MaybePromise` and
+    `await` it. A call that did not ask for a promise throws, naming the option and what the call
+    was building — `queryFromInfo` the field, `getQuery` the connection's model or table, which is
+    as much as it is given — rather than returning one where the declared type said there was none:
+    a promise spread into a prisma or drizzle call is one unusable key and no type error. Two things
+    make the query async. One is a selection beneath the field, which is a property of the document
+    rather than of the call site. The other is the helper's own `select` or `query` callback: an
+    async one makes `getQuery` throw for a fully synchronous document, since it is the helper's own
+    callback that has not settled. The return type follows the option's literal type, so the default
+    keeps the plain query; a non-literal `boolean` gets the `MaybePromise`, since the call cannot
+    rely on the synchronous type either.
+  - A selection passed to drizzle's `query()` comes first, as it does for `queryFromInfo`: a
+    relation the document also plans with other arguments loads on its own.
+
+- 1358f71: Cursors and node ids round trip for the column types a key holds — numbers, strings, bigints,
+  dates, booleans, decimals and byte arrays — and compound cursors are encoded with per-part type tags so
+  every value in one keeps its type, including booleans in Prisma compound unique keys.
+
+  A compound cursor was a plain JSON array of the raw values in the prisma plugin. That throws
+  outright on a bigint, so a connection whose cursor is a `@@id` or `@@unique` containing a `BigInt`
+  column failed on every edge of every page; it wrote a `DateTime` out as an ISO string that came
+  back as a string, so the millisecond a row was created on was gone by the time the value reached
+  the query; and it wrote a `Bytes` column out as the JSON form of a buffer. Every part now carries
+  its own type tag. The drizzle plugin already tagged each part; what changes there is that a
+  compound cursor can hold bytes, a decimal and a JSON value there too.
+
+  A null cursor value now round trips in both plugins. The prisma plugin threw "Unsupported cursor
+  type object" while building the edge of a single value cursor, so a connection whose cursor names a
+  nullable column failed the whole page rather than only a request that paged from that row; `Json`
+  and `Bytes` values could not be written there either. The drizzle plugin wrote the string `null`,
+  which is not a value the cursor format can hold, so the cursor came back as "Invalid cursor" the
+  moment a client paged from that edge. Both now write a null as a tagged chunk and read it back as
+  `null`.
+
+  Cursors issued before this release still parse, in both plugins, to the same values they always
+  did. That legacy handling is deprecated and is removed in the next major. The `GPC:` and `DC:`
+  prefixes are unchanged, so only compound cursors are written differently; a single value cursor is
+  byte for byte what it was, apart from the null that neither plugin could write usefully. A single
+  value string cursor also no longer loses everything after the first newline it contains.
+
+  Other values that did not survive a round trip:
+  - A cursor value on a float, real or `Decimal` column, in both plugins. It was read back with
+    `parseInt`, which truncates `1.75` to `1` and reads a large number's `1e+21` as `1`, so the page
+    resumed from the wrong row.
+  - A prisma node id built from a `DateTime` field. It was serialized with `String(date)`, whose
+    format carries no milliseconds, so `findUnique` looked for a row a fraction of a second before
+    the real one and the node resolved as not found. Ids already handed out in the old format still
+    parse. A drizzle node id built from a date column had the mirror problem on the way back: `new
+Date` was given the epoch milliseconds the serializer writes as a string, which is not a format
+    it parses, so every such id resolved as an Invalid Date.
+  - A prisma node id built from a `Bytes` field. The base64 case was spelled `Byte`, which is not
+    what the datamodel calls the type, so the value was serialized as `String(bytes)` (the bytes
+    separated by commas) and parsed back as that string.
+  - A drizzle node id built by `builder.drizzleNode` with an `id.column` that is not the table's
+    primary key. The id was parsed back into a record keyed by the database column name, while the
+    serializer wrote, and the model loader reads, the typescript name; the loaded rows were then
+    matched back to the ids they were loaded for by primary key, which such a record does not carry.
+    Such a node now loads.
+  - A prisma compound node id (a `@@id` or `@@unique` used as `id.field`). It passed the field's
+    `kind` (always `scalar`) where the type was wanted, so a `Json` or `Bytes` part was written as
+    `String(value)` and could not be read back. Each part is now serialized by its scalar type, as
+    the parser already read them.
+
+  Two paging fixes that go with them:
+  - A connection asked for `last: 0` pages backwards in both plugins, as any other `last` does. `0`
+    was read as "no `last` given", so the query took rows forwards and `pageInfo` came back as
+    though the request had been a forward page: `hasNextPage` true and `hasPreviousPage` false on a
+    non-empty connection.
+  - A prisma compound cursor carrying the wrong number of values is rejected with a validation error
+    naming both counts, rather than building a query with a missing key (or silently dropping an
+    extra one) for prisma to reject. The drizzle plugin already checked this.
+
+  The tagging itself lives in `@pothos/core` (`encodeCursorChunk`, `encodeCursorTuple` and
+  `decodeCursorChunk`, alongside `encodeBase64Bytes` / `decodeBase64Bytes`) rather than being written
+  out once per plugin.
+
+- 1358f71: `t.relatedCount` and a `t.relatedConnection`'s `totalCount` now count the rows the connection pages
+  over. The parent-to-child SQL is built by drizzle's `relationToSQL` instead of being reassembled
+  from the relation's source and target columns, so the plugin and the relational query builder agree
+  on what "related" means.
+  - New builder option `drizzle.filterConnectionTotalCount` (default `true`): `totalCount` on a
+    `t.relatedConnection` applies the `where` returned by the field's `query`, so it counts the same
+    rows the connection paginates. Previously it reported the unfiltered count. Set the option to
+    `false` to keep counting every related row regardless of the filter, matching the prisma plugin's
+    option.
+  - A count over a many-to-many relation (one defined with `.through(...)`) counts related rows. The
+    junction table was left out, which compared the parent's key to the target's key instead: the
+    count came back as the number of target rows that happen to share an id with the parent, with no
+    error. The same correction applies to the filter `t.relatedField` hands to its `select` callback.
+  - A count honours a `where` declared on the relation itself, and the `where` a relation without
+    `from`/`to` inherits from the relation it reverses. Both were dropped, so the count included rows
+    the connection does not page over.
+  - The count reads the relation's source column off the parent by its typescript name, as every
+    other read of a row does. A table declaring `postId: integer('id')` carries it as `postId`, so
+    the count bound `undefined` and the driver rejected the query with "undefined cannot be passed as
+    argument to the database".
+  - `totalCount` and the query that pages the relation are now the same predicate: the count selects
+    from what the page query selects from and repeats its `where` clause, minus the limit, the
+    ordering and the keyset clauses.
+  - A count over a many-to-many relation joins the junction table by its index. `t.relatedCount`
+    counts a related row once however many junction rows reach it; a `t.relatedConnection`'s
+    `totalCount` counts rows the way the page query returns them, so the two report different
+    totals for a relation whose junction holds duplicate rows.
+  - A `drizzleConnection` declared `nullable: false` no longer queries rows for a document that
+    selects only `totalCount`. The check that recognises a count-only selection read the field's
+    return type without unwrapping it, so a non-null connection never matched and always loaded a
+    page of rows it then discarded. Nullable connections were already unaffected.
+
+- 1358f71: When a field's data is missing from the parent row it falls back to the model loader, which
+  reloads the row. That fallback now loads the right row, and loads more of it.
+  - The loaded row also carries the parent type's type-level `select`/`include` (prisma) or
+    type-level columns, relations and extras (drizzle), since that row replaces the parent the
+    resolver sees. Fallback queries therefore load more than before: the parent type's own columns
+    and its type-level relations. The field's own selection always wins — a type-level relation
+    whose arguments conflict with it is left out of that query — and when a prisma type-level
+    `_count` has several counts and one conflicts with the count a field selects for itself, the
+    loader keeps the other counts instead of dropping `_count` altogether.
+  - Drizzle: `t.relation`, `t.relatedCount` and `t.relatedConnection` fall back to the loader when
+    their data is missing from the parent row — most often because the row was returned by a
+    resolver that fetched it itself. Previously the loader was reached only when the field had no
+    loader mapping at all, and a row without the relation on it resolved as `undefined` or as an
+    empty page. `t.relatedConnection` decides what it has by reading its selection at resolve time
+    the same way the planner does, through fragments, `@skip`/`@include`, and a wrapping type, so a
+    row that carries the relation rows but not the count reloads rather than reporting `totalCount`
+    as `undefined`.
+  - Prisma: a model load that failed before its query was sent hung the request instead of returning
+    an error. The loader batches the rows it reloads and sends one query per row on the next tick. A
+    throw while building any of those queries, most reachably `findUnique: null`, which throws
+    `Missing findUnique for <type>`, abandoned every row it had not reached yet, so those fields
+    never resolved and never errored and the request never finished. Every row in the batch now
+    rejects with that error, matching what `@pothos/plugin-drizzle` already did.
+  - Prisma: the per-row fallback plan is cached. When a relation or related connection has a
+    user-supplied `resolve` and the planned query did not load it, the field's resolver plans the
+    field for itself; that planning ran again for every row of the list its parent came from. It is
+    now cached per `Type@path` per request, the way `ModelLoader` already caches its own entry point
+    and under the same key: the plan is settled once and played per call, so each row still gets a
+    query of its own and still records the same mappings under the same keys. On a document with 161
+    such resolutions the planning cost falls from about 405µs to about 68µs per request, with the
+    queries issued and the response unchanged.
+
+- 220a1b1: The selection API is now typed to match what it actually does.
+
+  Two changes can stop existing code from compiling. In the first case that code already failed at
+  runtime:
+  - `t.relationCount` (prisma) and `t.relatedCount` (drizzle) accept only list relations. Calling
+    either on a to-one relation built a count query prisma or drizzle rejected.
+  - `LoaderMappings` (prisma) is a deprecated alias of the shared `Mappings` type. The record it
+    names is internal to the plugin, and the plugin no longer builds the `{ field, type, mappings,
+indirectPath }` entries the old declaration described, so code that reads those keys off it no
+    longer compiles. Nothing in the plugin's API hands one to a schema.
+
+  New and widened types:
+  - `queryFromInfo` (prisma) is typed by what was passed: `{ include }` for a given `include`, and
+    `{ select }` keeping a given select's own type, so the selection round trips. Spread the result
+    into a prisma call and read the selected columns and relations back off the rows, with types. A
+    call with neither now returns `{ select?, include? }`, whichever of the two the type's mode
+    produced, rather than a `select` key that a type in include mode does not return.
+  - `nestedSelection` returns the relation query for the field's model or table (prisma:
+    `PrismaRelationQuery<Model>`; drizzle: the table's `DBQueryConfig`, a `many` config for a list
+    field), keeping the keys it was given as given, so a `select`/`columns` in them still narrows
+    the parent shape. Its argument is typed by that query when the field's type names a model, so
+    literals such as `select: { title: true }` are kept, and it accepts a callback or a promise as
+    the runtime always did. It used to return its argument's type.
+  - One `PathSegment` type (`string | { name: string; type?: string }`) for `queryFromInfo` paths
+    and `nestedSelection` paths, exported from `@pothos/selection-mapper` and re-exported by both
+    plugins, together with `IndirectPathSegment` (its `{ name, type? }` form) and `IndirectInclude`,
+    whose `path`/`paths` still take only `IndirectPathSegment` objects. A `{ name, type }` segment
+    given to `nestedSelection` now pins the implementation the field is found under at runtime, as
+    it already did for `queryFromInfo`.
+  - The `query` a `t.relation` fallback `resolve` receives (prisma) — the relation's own arguments
+    (`where`, `orderBy`, `take`, ...) alongside the planned `select`/`include` — has a name:
+    `QueryFromRelation`.
+  - `t.relatedField` (drizzle) accepts the ordinary field options (`deprecationReason`,
+    `extensions`, `authScopes`, ...), and forwards every one of them to the field it builds, merging
+    `extensions` rather than dropping them; `t.relatedCount`, which is built on it, gets the same.
+    Its `resolve` may be async and receives the resolve `info`. Exported as
+    `RelatedSelectionFieldOptions`.
+  - A query whose `select` key is optional but names columns (`{ select?: { email: true } }`)
+    narrows the row to those columns, rather than to no columns at all as `ShapeFromSelection` did
+    before: the columns are on the row whether or not the `select` is applied, since a row loaded
+    without a `select` carries every column.
+
+  The docs for both plugins now cover how a field gets its data, and when it costs a query of its
+  own. Also newly documented: prisma `findUnique: null`, `onNull`, `nestedSelection`'s typed
+  segments and third argument, `skipDeferredFragments`,
+  `prismaFieldWithInput`/`drizzleFieldWithInput`, drizzle
+  `drizzleInterface`/`drizzleInterfaceField(s)`, `t.variant(Ref, { select })`, connection size
+  options, the drizzle fallback loader, and the error two variants raise when their selections
+  conflict.
+
+- 220a1b1: The prisma and drizzle plugins now plan their queries with one shared selection planner, the new
+  package `@pothos/selection-mapper`. Each plugin keeps its own adapter for its query format; the
+  walk from a GraphQL selection to a query is the same code in both.
+
+  The plugins' public API is unchanged: only internal helpers moved (each plugin's `map-query`,
+  `selections`, `loader-map` and `deep-equal`), and the loader mapping keys (internal) now carry the
+  full field path. Behaviour that differed between the two plugins for no reason now agrees. In
+  particular, prisma walks fragments the way drizzle did: fields under a fragment on an interface the
+  type implements, including interfaces not backed by a model, are planned for the type instead of
+  falling back to a query per row.
+
+  The package is an implementation detail of the two plugins. Its exports exist for them and change
+  with them: it is not a supported public API, and its versions carry no compatibility promise of
+  their own.
+
+- 1358f71: Entering another type of the same model through a fragment (a variant under a model interface or
+  union) now merges that type's type-level selection, so resolvers relying on it find their data in
+  the same query. This can change the shape of queries a schema with variants issues, and it adds one
+  validation error.
+  - A variant without a type-level `select` switches the query to include mode (prisma) or every
+    column (drizzle).
+  - If the two types' type-level selections disagree on a relation's arguments, a validation error
+    names both types and the relation; move the relation arguments to a field-level `select` on one
+    of the types. The error is raised whichever route brought the two types together: a fragment
+    entering a variant, or two indirect-include path matches landing on different types of the one
+    model. Relation arguments are compared structurally, so equal arguments written
+    separately are fine. In drizzle, two types defining the same `extras` key with different
+    functions are rejected the same way, naming the extra: an extra shared across variants must be
+    the same function reference.
+  - The type-level selections of every variant entered at a selection set are merged before the
+    fields there are planned, so the result no longer depends on document order: a field-level
+    `select` whose relation arguments conflict with a variant's type-level selection falls back to
+    its own query whichever comes first, and only two type-level selections can conflict.
+  - A walk on an object type is a walk on rows of that type: the field's own type, the `typeName`
+    given to `queryFromInfo`, the type pinned by `nestedSelection`, or a node load. A fragment on
+    another object type of the same model cannot apply to those rows, so it is not entered; it is
+    entered while walking an interface, where rows may resolve to it. Prisma previously entered it
+    whenever the field's declared type was abstract, planning a sibling variant's selection for
+    rows that could never be it.
+  - A fragment that does not apply to the type still lets a nested fragment narrow back to it.
+    Drizzle previously stopped at the outer fragment and discarded what was nested inside it.
+
+### Patch Changes
+
+- 220a1b1: Drizzle plugin fixes.
+  - `t.relatedConnection` calls its `query` callback with the same `pathInfo` while resolving as
+    while planning, so a query that branches on `pathInfo` pages the rows it selected.
+  - `t.relatedConnection` no longer accepts a `resolve` option. It was silently discarded, so the
+    resolver never ran.
+  - `t.relation` evaluates a `query` callback once, with `pathInfo`. Previously it was called a
+    second time without `pathInfo`.
+  - `drizzleConnectionHelpers(builder, table)` can now be called without the options argument.
+    Previously it threw.
+  - `builder.drizzleInterfaceField(s)` given a table name resolves it to the interface registered for
+    that table, as the docs describe, instead of the object type.
+  - `pathInfo.path` handed to relation `query` callbacks now starts with the root field for queries
+    planned through `path`/`paths`, as it already did for queries planned directly for a field.
+  - `pathInfo.segments[].isList` is `true` for non-null list fields (`[Post!]!`). Previously only
+    nullable lists were reported as lists.
+  - The error for a `drizzleObject` and a `drizzleInterface` it implements being built on different
+    tables names drizzle rather than prisma. It read "must be based on the same prisma model as any
+    DrizzleInterfaces they extend"; it now says "the same drizzle table".
+  - Custom client adapters must now provide `select()` and its SQL builder chain, plus
+    `query.<table>.findMany()`, alongside relation metadata and `$count()`. Forward SQL construction
+    synchronously; do not wrap `select` in a promise or Effect. Full Drizzle clients already
+    provide these methods.
+  - Requires Drizzle `1.0.0-rc.2` or newer for the relation SQL API used by counts and predicates.
+  - Custom connection fields retain their rows when `totalCount` is selected.
+  - Connection helper `resolve()` returns a maybe-promise for async-enabled schemas; await it
+    before inspecting or spreading connection fields.
+  - Nested query callbacks passed to `nestedSelection` now receive `PathInfo` as their third
+    argument, after the field arguments and context.
+
+- 1358f71: Fix how selections are found through fragments, paths and wrapper types, in both plugins. Each of
+  these left a field to load its own data per row, or planned the wrong selection for it.
+  - A path segment selected behind a fragment that narrows an interface or union
+    (`... on Impl { field }`), or a fragment on an interface that overlaps the list type, is now
+    found without needing to pin the type on the segment. Previously the field was skipped and the
+    plugin fell back to a separate query per row. Drizzle crashed on a fragment spread that narrows
+    an interface.
+  - When the same field is selected in several places (multiple fragments, several `paths`, or the
+    same response key under several fragments), every selection is merged into one query. Previously
+    `queryFromInfo` kept only the last one, and a field loaded through the fallback planned only its
+    first occurrence.
+  - A relation selected under both `nodes` and `edges { node }` of a connection is answered from the
+    loaded rows in both places. The loader mappings used to be keyed by field alias alone, so the
+    second path fell back to a query of its own.
+  - A `nestedSelection` path into a field selected through a fragment on the same type
+    (`... on Type { field }` or `...Fragment`) now merges the nested selection instead of producing
+    an empty selection.
+  - Fragment spreads and inline fragments left out by `@skip` / `@include` are no longer planned, so
+    they neither enter a variant nor add fields to the query.
+  - `queryFromInfo`'s `path` and `paths` accept `{ name, type }` segments alongside strings, in the
+    types and the docs as well as at runtime. A segment `type` pins the fragment type condition the
+    field must be found under, which is useful when several implementations share a field name. The
+    pin is applied wherever the field is looked for, not only where a fragment was crossed to reach
+    it: a field of the same name selected directly on a level that is not the pinned type is no
+    longer taken as the segment's, so the match no longer carries that level's return type instead
+    of the pinned type's. A segment `type` that does not exist in the schema is a validation error,
+    rather than a `TypeError` (prisma) or a path that silently matches nothing (drizzle).
+  - When several implementations share a field name, matches whose field returns a different model or
+    table than the one the query is being built for are ignored, so their selections are never merged
+    into the wrong query. Variants of the target model are walked with their own fields.
+  - A type carrying `pothosIndirectInclude` with `paths` only contributes its own selection when it
+    is backed by the model being queried. A plain wrapper, or a wrapper backed by another model, no
+    longer merges its selection into the target model's query.
+  - `queryFromInfo` with `path`/`paths` that select nothing now returns the `select`/`include`
+    (prisma) or `select` (drizzle) it was given, or `{}`, instead of an empty `select` (which prisma
+    rejects) or a query that dropped the caller's selection.
+  - The selection lookup handed to a field's `select` callback (the fourth argument) now looks
+    through wrappers on the field's return type and returns the first match. A `relatedConnection`
+    with `totalCount: true` wrapped by `@pothos/plugin-errors` now selects the count, and its
+    totalCount-only handling agrees with what was planned.
+  - A field `select` callback returning `false`, `null` or `undefined` is treated as selecting
+    nothing: the field is not answered from the parent row and loads its own data when resolved.
+    `false` used to switch the query to include mode and still record the field as loaded, so it
+    resolved from a row that held nothing for it; `null` and `undefined` threw.
+  - Relation arguments compare the same way whichever of two selections was planned first. The
+    structural equality both plugins use decided whether a value was an array, or a boxed value such
+    as a `Date`, by looking only at its left-hand argument: `{}` and `[]` compared equal one way
+    round and unequal the other, and so did `{}` and a `Date`. That helper answers whether two
+    selections of one relation ask for the same arguments, so an answer that depended on argument
+    order could let an incompatible selection merge and replace the arguments already planned, or
+    refuse a compatible one and issue a separate query for the field. Both guards now read both
+    sides. `orderBy` is the argument where this is easiest to hit, since prisma accepts an object and
+    an array there for the same meaning.
+  - @pothos/selection-mapper@0.1.0
+
 ## 0.18.1
 
 ### Patch Changes