Update review · September 1, 2026
Cursor-Based Pagination for Multi-Valued Attributes in SCIM (-01 update)
draft-kushwaha-scim-attr-cursor-pagination-01 — Saurabh Kushwaha (Oracle), submitted 30 August 2026. Same single author. Intended status changed from Informational to Standards Track, and the header now says Updates: 7643, 7644 (if approved).
The -00 review said this document was about a quarter of the way to being a serious proposal, that it engaged with none of the three existing approaches to the same problem, and that it had a data-loss bug at its center. Four pages, one normative reference, no discovery, no error codes.
The -01 is 48 pages with seven normative references, a 12-page comparison against prior art, IANA registrations, and an appendix of SCIM schema definitions. Every substantive item from the previous review is addressed. This is the most complete turnaround I've seen on a -00 review.
What Got Fixed
Point by point against the -00 feedback:
RFC 9865 is now a normative reference with a dedicated comparison section. Section 16.1 states which of its design properties this document reuses: opaque forward cursors, provider-selected defaults, maximum page sizes, cursor expiry, per-page authorization, and no fallback after an invalid cursor. The draft keeps distinct scimType names (invalidAttributeCursor, expiredAttributeCursor, invalidAttributeCount) rather than reusing 9865's, which is defensible since the scope differs.
Phil Hunt's multi-value work is cited and argued against on the merits. Section 16.2 goes further than citing it. It records that the May 2022 call for adoption of draft-hunt-scim-mv-filtering failed for lack of responses, that the one substantive response opposed adoption, and what the grounds were. Then it names the two properties the qualifier approach had that this document deliberately gives up: backward compatibility and statelessness. Section 16.2.1 is a new subsection on Active Directory range retrieval and the ordering defect that index qualifiers inherit when another client modifies the attribute mid-traversal. That's a real argument, sourced from a mechanism that's been deployed for twenty years.
Danny Zollner's approach gets a section, and the draft doesn't claim to replace it. Section 16.4.1 says top-level relationship resources solve relationship management, this document solves bounded retrieval, and the two are orthogonal. The argument is that a provider exposing /GroupMembers still has to answer GET /Groups/{id}, because Group.members is still returned: default in RFC 7643 no matter what other endpoints exist. That's the correct framing and it's the strongest thing in the document.
The single-cursor-many-attributes problem is solved, differently than Hunt solved it. The client never names an attribute in a cursor. The provider hands out a per-attribute firstPage link. A collection search defers every protected attribute independently, each with its own link. At most one attribute gets hydrated per response. So a cursor is always bound to exactly one attribute on exactly one parent, and the ambiguity disappears instead of getting encoded.
The data-loss bug has a section. Section 12 says a page or deferred representation MUST NOT be used as the source for a full-resource PUT or a whole-attribute PATCH, cites RFC 7644's rule that PUT can clear omitted read-write attributes, and adds a partialAttributeReplacement error. Section 4 adds the rule that matters more: an omitted attribute MUST NOT be represented as an empty array or null, because either one is indistinguishable from complete state.
Discovery exists. A ServiceProviderConfig extension with defaultMode, defaultPageSize, maxPageSize, cursorTimeout, totalResults support level, and a resourceTypes list naming which attribute paths are protected per resource type. Plus the rule that a client MUST NOT read an omitted numeric limit as unlimited.
The metadata has a real schema. membersPagination and its implied rolesPagination are gone. Response metadata is an attributePagination array inside urn:ietf:params:scim:schemas:extension:attributePagination:2.0:Resource, with full schema definitions in Appendix A and IANA requests in Section 17.
Ordering, errors, defaults, and the attributes interaction are all specified. Section 10 mandates a deterministic total order with a unique tie-breaker appended when the sort value isn't unique. Section 13 defines seven scimType keywords. Section 8.1 spells out the interaction with attributes, excludedAttributes, returned: always, and authorization.
hasMore is gone, with a sentence explaining that nextCursor presence is authoritative and a second signal could drift out of sync. itemsPerPage is redefined as the count actually returned. totalResults is optional with a discovery field and a section saying providers SHOULD omit it when counting is expensive.
The boilerplate problems are fixed. This went through xml2rfc. Table of contents, BCP 14 section, split normative and informative references, page numbers.
What's New
The document doesn't just answer the -00 feedback. It grew a second half.
Deferred collection retrieval is the new headline mechanism. A collection search returns parent resources with the protected attributes absent, each carrying a provider-generated firstPage link. No count query, no identifier list, no hydration while building the parent page. The state field distinguishes deferred from page. This is a bigger deal than the cursor half, because the collection fan-out case is where a sync client actually falls over: 100 groups in a ListResponse, each triggering an unbounded membership query.
Automatic vs explicit modes. Explicit mode only paginates when the client sends attributeCount or attributeCursor. Automatic deferred mode lets the provider defer and apply default page sizes even when the client sends nothing. Section 16.2 explains why automatic mode has to exist: a mechanism that only protects the provider when the client cooperates does not protect the provider. An unmodified client issuing a bare GET /Groups/{id} against a two-million-member group has no way to know it should ask for less.
An RFC 7240 preference. Prefer: scim-attribute-pagination=deferred for in-band client opt-in, with Preference-Applied on the response, and an IANA registration request.
Look-ahead row rules. Retrieve effectivePageSize + 1 candidate rows from the query that discovers the values, drop the extra row before hydration, and issue nextCursor only when that row proved another page exists. Section 8.5 adds the rule for reference-derived attributes: paginate at the relationship query, never collect every referenced identifier and then paginate the hydrated result. There's even a warning that a 1000-value page plus look-ahead must not become a 1001-expression IN predicate.
Budget-driven early termination. A provider MAY stop a page short of the effective page size when a byte, time, or hydration budget is hit, and MUST return a cursor whose boundary is the last value actually returned. It MUST NOT return a cursor that advances past values skipped for budget reasons.
Optional attributeVersion for providers that can offer stronger change detection, explicitly not required to equal the parent's meta.version.
Rebuttals of two mailing list suggestions. Section 16.5 explains why paginating the serialized JSON object doesn't work: no attribute semantics, so it can't honor returned: always or distinguish withheld from empty, and the boundary isn't addressable. Section 16.6 explains why chunked transfer encoding isn't an answer: it converts a large allocation into a long-running one without bounding any of the work.
Section 16.8 concedes the 2022 objection. Matt Peterson's argument was that group membership use cases are already satisfiable with GET /Users?filter=groups.value eq "..." and ordinary resource pagination, no extension needed. The draft accepts that analysis, accepts that the base schemas contain few attributes big enough to warrant this, and then names three limits: inverse filtering answers a different question than "what does the provider return for GET /Groups/{id}", User.groups filtering support is uneven and undiscoverable, and derived assignment attributes often have no counterpart resource to filter over. Peterson wrote RFC 9865, so this is the draft arguing with its own primary normative reference's author, and it does it fairly.
The acknowledgements section ends with "Acknowledgement does not imply endorsement. Several of the contributors named above have expressed reservations about the approach taken in this document, and Section 16 records those positions." That's the right way to handle it.
What Got Removed
hasMore. The membersPagination convention-generated naming. The one-sentence "Rationale" section. The "Future Work" section, replaced by an explicit list of what this version does not define. The Informational status. And the -00 claim that "existing SCIM clients remain unaffected," which the -01 explicitly disclaims: "This specification does not claim that automatic deferred mode is wire-compatible with clients that require a complete default attribute."
Implementation guidance moved from the body into non-normative Appendix C, which is where SQL examples belong.
What's Still Open
Automatic deferred mode is the crux, and its negotiation isn't interoperable. The document updates RFC 7644 §3.9 for "parties for which deferred mode has been established." The only in-band, interoperable way to establish it is the Prefer header, and Section 4.2 says the absence of that header doesn't prevent deferral. Everything else is "a staged tenant or client profile, an out-of-band capability agreement, or another deployment-specific negotiation mechanism." So the condition that gates a normative change to RFC 7644 is out of scope of the document that makes the change. Two conformant implementations can't agree on when §3.9 applies, and you can't write an interop test for it. The draft is honest about this rather than papering over it, which is better than the alternative, but it's the thing the working group is going to spend its time on.
attributePaginationRequired has no defined client remedy. Section 4.2 says a provider that can't safely return an unbounded attribute SHOULD fail closed with this error. Nothing says what the client does next. Retry with attributeCount? Send the Prefer header? Add excludedAttributes=members? "Deferred behavior has not been established for the client" is not something a client can act on by itself.
cursor: false in discovery is undefined. Both supported and cursor are REQUIRED booleans. If the extension is present at all, supported: true carries no information. And a provider advertising cursor: false is offering deferral with a firstPage link that returns one page and no way to reach the rest, which is truncation with better branding. Either say what that combination means or drop the field.
maxHydratedAttributesPerResponse is an integer with exactly one legal value. The table says "This version defines a maximum of one." Make it a constant and remove the field, or define what happens at 2, including how two independent cursors coexist in one response.
Mutation safety enforcement is best-effort by design. partialAttributeReplacement fires when the request carries the extension metadata, and Section 12 admits a provider can't tell how a client built a request once the client strips it. The absent-not-empty rule in Section 4 is the actual protection and it's a good one. But the enforcement story falls back to "deployment negotiation and client-profile controls remain necessary," which is the same escape hatch as the compatibility model.
POST /.search still isn't mentioned. Section 8.2 requires rejecting attribute-level parameters on collection endpoints, and /.search is a collection endpoint, so the rule arguably reaches it. But /.search carries its parameters in a JSON body rather than the query string, and the document never addresses that.
Section 16.9 is an open issue the author raised himself, and it's the right one. As specified, traversing a protected attribute takes one request per parent. A client enumerating membership across 10,000 groups makes at least 10,000 requests. The author notes this forecloses a root-level form like GET /Groups?attributes=members&attributeCursor=<opaque>, which is a prerequisite for delta query over relationship sets, sketches how a composite (parent identifier, value identifier) ordering could express it, and puts four unresolved questions to the working group including whether the work belongs here or in the relationship-resource draft. That last question deserves a direct answer on the list.
Five stale cross-references, all in the comparison sections. Section 15 points at "Section 14.3" for work budgets, which moved to Appendix C.3. Section 16.2 points at "Section 2" for unbounded work when it means 1.2, and "Section 11" for cursor construction when it means 10. Section 16.2.1 points at "Section 7" for deterministic total ordering when it means 10. Section 16.4.2 points at "Section 6" for firstPage when it means 7. Mechanical, but Section 16 is the part reviewers will read hardest.
The Zollner reference is wrong. [I-D.zollner-scim-groupmember] cites "draft-zollner-scim-groupmember-resource-extension." The draft is draft-zollner-scim-group-members, currently at -01.
Background of the Author
Unchanged. No new authors. Saurabh Kushwaha (Oracle) still has zero RFCs, zero expired drafts, and four active drafts, all SCIM, all submitted in 2026.
The -00 review flagged the volume as the thing to watch and called this the thinnest of the four. That's no longer true. This is now the most developed document of the set, and the revision cycle demonstrates the author reads feedback carefully and acts on it. The -00 review said the draft "reads like it was written before anyone looked at the existing work." Someone has now looked at the existing work more thoroughly than most adopted drafts manage.
Suggested Feedback
The document is no longer the problem. The design question is.
Ask the direct question about automatic deferred mode: can a normative change to RFC 7644 §3.9 be gated on a negotiation mechanism the document doesn't define? Either the Prefer header becomes the required trigger, in which case explicit mode is all you have and the provider-protection argument in Section 16.2 goes unanswered, or the document has to say something interoperable about how a compatibility profile gets established. This is the fork the working group has to pick.
Answer Section 16.9. If root-level traversal belongs in the relationship-resource work, say so, and this document shrinks to the single-parent case. If it belongs here, the composite ordering needs specifying now, because it changes the cursor contract.
Say whether the Section 16.4.1 orthogonality argument holds. The claim is that Group.members being returned: default in RFC 7643 means every provider needs bounded retrieval regardless of what other endpoints it offers. The counter-argument is that the revision of RFC 7643/7644 the charter authorizes could change that characteristic and make the whole problem go away. Somebody should say that out loud.
Ask about attributePaginationRequired recovery, and pin down cursor: false and maxHydratedAttributesPerResponse.
The -00 recommendation was to bring this to the list as input to the "Multi-Value Query Filtering and Paging" charter item rather than as a fourth parallel draft. That recommendation stands, but for the opposite reason now. It's no longer too thin to be input. Section 16 is a genuinely useful survey of every approach the working group has considered since 2019, including the ones that lost, and that survey is worth having in front of the group whether or not this particular mechanism is what gets adopted.