Review · November 16, 2026
SCIM Group Membership Resource Type
The Problem
A group with a million members does not fit in one HTTP response.
RFC 7643 stores membership as a multi-valued members attribute on the Group. SCIM paginates resources, not values inside a resource. The draft puts a million-member Group at roughly 200MB (Section 2). Some service providers have stopped returning members at all. Some clients exclude it on every request, so they can push changes but never read back the real state.
The fix is to make each membership its own resource at /GroupMemberships. Then pagination and filtering work the way they already do for Users.
Is This a New Draft?
No. It is a rename. Datatracker lists this draft as replacing draft-zollner-scim-group-members, which we reviewed in August. The -00 is misleading. This is effectively the third revision.
The resource is now GroupMembership instead of GroupMember. The endpoint is /GroupMemberships. The design is the same, and most of the text around it was rewritten.
How It Fits In
The SCIM charter lists "Multi-Value Query Filtering and Paging" as part of the RFC 7643/7644 revision. This draft does not page a multi-valued attribute. It moves the data out of the attribute. That makes it an alternative to the chartered item, not an instance of it. The session at IETF 127 has a rechartering discussion on the agenda, so this is the time to settle it.
The competing approach is draft-kushwaha-scim-attr-cursor-pagination, which adds new query parameters to page inside a resource.
This draft adds one resource type, one Group schema extension, and two schema URNs. No new roles. No new query syntax. It depends on RFC 7643, RFC 7644, and RFC 9865 for cursors.
Summary
A GroupMembership is one link: a group reference, a member reference, plus id and meta (Section 4.1). Only direct memberships count. One resource per group and member pair.
- Add a member with
POST /GroupMemberships. A duplicate gets 409 withuniqueness(Section 6.1). - Remove one with
DELETE /GroupMemberships/{id}(Section 6.3). - There is nothing to PATCH or PUT. Both references are immutable (Section 6).
- Servers MUST support
eqfilters ongroup.valueandmember.value, and the two combined withand(Section 6.2.2). - Servers MUST page with index or cursor pagination and SHOULD support cursors (Section 6.2.1).
The Group extension is one boolean, membersEmbedded (Section 5.1). True is a guarantee that members holds the full membership. False means no guarantee, so go to /GroupMemberships.
A server that keeps members writable MUST keep both views consistent, and deleting a Group or a member MUST delete its memberships (Section 6).
What the Rename Fixed
Most of the August review's complaints are handled.
- Delete needed a lookup that might not work. The compound filter is now mandatory, so a client can always find the
id. Section 6.3 also points to PATCH on the Group as a path that needs noid. - No ordering. Servers MUST return a stable order across the pages of a query (Section 6.2.1).
- Unfiltered
GETreturned everything. Servers MAY reject it with 400tooMany, and SHOULD cap page size. - Three policy values.
inline,externalandhybridare gone. One boolean replaces them. ref,memberCount,allowedMemberTypes. All removed.- "Transactional integrity." Replaced with a consistency requirement and two concrete examples.
- Security and IANA nested under the schema section. Both are top-level now. Security Considerations grew to four specific rules, including not leaking existence through errors or
totalResults(Section 8).
What Is Still Open
A PUT can still wipe a group. This was the first item in August and the draft says nothing about it. A server MAY omit members (Section 5). A client reads the Group, changes displayName, and PUTs it back without members. RFC 7644 Section 3.5.1 lets the server treat the omitted attribute as cleared. The new boolean makes it slightly worse. With false, members "may in fact be full, partial, omitted, or empty", so a client can write back a partial list.
A POST that times out is recoverable, but only by lookup. The 409 still does not carry the existing resource's id or location. The client has to run the compound filter to get it.
Read-only support is still signaled through mutability. A read-only server sets every attribute to readOnly in /Schemas (Section 4.1). The draft adds 501 for POST and DELETE, so a client can also find out by failing.
bulkId is still not shown. Creating a User and adding it to a group in one bulk request is the common case. Section 6.4 only uses existing ids.
Three new things to check:
- Pagination is "at least one of" index or cursor. A cursor-only server and an index-only client cannot talk.
/ServiceProviderConfigtells the client, but it cannot fix it. member.typeis not required in the schema. Filtering on it is a MUST when a server allows more than one member type. The draft should saytypeMUST be populated in that case.tooManyis defined in RFC 7644 for a filter that yields too many results. Here it is returned for a request with no filter.
Writing Quality
Clean and compact. Standard IETF structure, BCP 14 language, complete registration templates. About half of the 24 pages are JSON.
Small issues:
- The Introduction still reads "protocol RFC7643 [RFC7644]". It names one RFC and cites the other.
- The abstract lists sorting as a missing feature. The body never mentions
sortBy. - There are no informative references. The draft does not cite the competing pagination draft, or the draft it replaces.
The Acknowledgments thank three reviewers from the SCIM list and the repository. The amount of change since -01 suggests that review was taken seriously.
Author
From his Datatracker profile, Danny Zollner (Okta) has 1 RFC, RFC 9865 on SCIM cursor pagination. He has 4 active drafts, all SCIM: this one, draft-zollner-scim-interop-profile, draft-schreiber-scim-ipsie-profile and draft-wzdk-scim-agent-resource. He has 4 expired drafts, including the working group document draft-ietf-scim-roles-entitlements. Not a new author.
Questions for the Working Group
So what should the room decide in 20 minutes?
- Does the group want a membership resource, in-resource pagination, or both? Pick on purpose. The rechartering discussion should name which one the charter means.
- If this one, is it ready for an adoption call? The design is stable and the August gaps are mostly closed.
- What does a server do with
membersin a PUT when it does not return the full list? This needs an answer before adoption. - Should cursor pagination be a MUST for this resource?
- Should the 409 return the existing membership's location?
- Does anyone have an implementation? The draft lists none.