Release Cadence Adjustment: SLURP Model

Note

This document is originally copied from the TC resolution 2022-02-10 Release Cadence Adjustment

History

OpenStack has historically used a six month release cycle cadence for the projects which participate in the coordinated release. Further, upgrades were tested and supported between two adjacent coordinated releases only, requiring deployers and distributions to either upgrade every six months to stay current, or perform Fast Forward Upgrades (FFUs) to move between non-adjacent releases at runtime. The latter is an activity enabled by testing the individual upgrade steps, and is not something we test specifically.

Challenges

Some deployers and distributions have indicated that six month upgrades are difficult, infeasible, or undesirable, especially in large environments where the process itself takes long enough that upgrades are constantly occurring. The FFU process can be laborious and also requires running parts of a release that may have never been deployed, productized, or tested in a given environment - purely because each release must be used stepwise during the operation.

A number of opinions have been expressed about changing the release cycle to either a slower (one year) or much slower (18 month) cadence to address these concerns. The lack of consensus around what that slower cadence should be makes it difficult to choose one that will be beneficial, as one cycle length may cause people on a slower cycle to need to wait much longer between upgrades. Further, community involvement in a very long release can be difficult when attrition, turnover, contract obligations and volunteer realities make slowing down unpalatable in many cases. The community already struggles to find candidates for six month (PTL) and one year (TC) duties. Further, for environments that do need to move quickly, adopt new features and deploy new technologies, double-digit months between landing a feature and having it testable and usable in production is too long.

Proposed Solution

It is very difficult to settle on any one change to the release cadence that will address all of the above problems and concerns. As such, the TC proposes an incremental change in release upgrade expectations to help improve the slow-moving deployer experience, without sacrificing the release-early-release-often goal.

The fundamental change comes to the expectation that upgrades are only supported between adjacent coordinated releases. The TC will designate major releases in a new arrangement, such that every other release will be considered to be a “SLURP (Skip Level Upgrade Release Process)” release. Upgrades will be supported between “SLURP” releases, in addition to between adjacent major releases (as they are today). Deployments wishing to stay on the six-month cycle will deploy every “SLURP” and “not-SLURP” release as they always have. Deployments wishing to move to a one year upgrade cycle will synchronize on a “SLURP” release, and then skip the following “not-SLURP” release, upgrading when the subsequent “SLURP” is released.

Our letter-based release naming scheme is about to wrap back around to A, so the proposal is that the “new A” release be the first one where we enforce this scheme. Y->A should be a “dress rehearsal” where we have the jobs enabled to help smoke out any issues, but where hard guarantees are not yet made.

Occasionally, individual releases are chosen by a large number of deployers and distributors by chance, which results in a larger than normal community of maintainers that keep the release “alive” in extended maintenance for longer. The expectation with this proposal is that this will amplify that effect by increasing the likelihood that “SLURP” releases will be chosen in this way and thus end up with more focus on those releases for long-term community support.

Details

  1. Testing: Just as we test and guarantee that upgrades are supported between adjacent releases today, we will also test and guarantee that upgrades between two “SLURP” releases are supported. Upgrades are tested for most projects today with grenade. A skip-level job will be maintained in the grenade repository that tests a normal configuration between the last two “SLURP” releases. The job will be updated on every new “SLURP” release, and there will always be a regular single-release grenade job testing between the previous release and current one, as we have today.

  2. Not-SLURP upgrades: Upgrades from “not-SLURP” to “not-SLURP” will not be tested nor required. On a given “not-SLURP” release, the only upgrade path will be to the following release (which would be a “SLURP”). This is unchanged from today.

  3. Intervals: Upgrades that span more than one “SLURP” cycle are not tested or required. For example to move from “SLURP A” to “SLURP E” will still require an FFU style arrangement, but where “SLURP C” is the only intermediate step required.

  4. Deprecations: Projects currently deprecate features and config for at least one cycle before removal. This change affects when that can happen, so that no required changes occur in a “not-SLURP” release which may be skipped. Effectively the same rules that we have today (both written and tribally-understood) apply to the new arrangement, with the exception that “cycle” refers to a “SLURP to SLURP” cycle and not a single pair of adjacent coordinated releases. Since the deprecation, waiting, and removal can only happen in “SLURP” releases, the result is also that the minimum length of time that things may be deprecated before removal will increase as well.

  5. Support: We will expect to support both the most recent “SLURP” release as well as the one prior. During a “note-SLURP” release, that would effectively be similar to what we support today, which is 18 months of “maintained” releases. See the example sequence below.

  6. Rolling Upgrades: This scheme does not necessarily dictate that live or rolling upgrades need to be supported between “SLURP” releases. Meaning RPC compatibility between N to N-1 guarantees can remain, resulting in deployments that are on a “SLURP to SLURP” release schedule requiring some downtime during an upgrade because components will be spanning more than two actual releases.

  7. Data migrations: Part of supporting “SLURP to SLURP” upgrades involves keeping a stable (read “compatible” not “unchanging”) database schema from “SLURP to SLURP”. This includes data migrations which need to do work in “SLURP” releases, and while they may do work in “not-SLURP” releases, the work done in “not-SLURP” releases cannot be mandatory. This can be solved (as it is today) by requiring operators to (force-)complete data migrations on a supported release before moving to one that drops compatibility. The “SLURP”, “not-SLURP” arrangement described in this document would require attention to those migrations to make sure they happen (automatically or manually) on the source “SLURP” before upgrading to the target “SLURP”, for example.

  8. Communication: We will use “SLURP” word to designate a SLURP release in release page, release notes page or any other place we want to communicate it. We can also use its full form “Skip Level Upgrade Release Process” if needed. A “not-SLURP” release will not be designated with anything and not having “SLURP” word is enough to communicate that this is not “SLURP” release. Also, the number schema in the release naming process will help all of us to relate which release is “SLURP”.

Example sequence

Assuming that A is the first release of this new arrangement, the following examples help demonstrate the support lifecycle expectation.

Release

Type

Supported

EM

A

SLURP

X,Y,Z

W

B

Y,Z,A

W,X

C

SLURP

A,B,C

W,X,Y,Z

D

A,B,C,D

X,Y,Z

E

SLURP

C,D,E

Y,Z,A,B

F

C,D,E,F

Z,A,B

G

SLURP

E,F,G

A,B,C

(EM releases are arbitrarily pruned in the above example for brevity, but no such change in how long they may be supported is made in this document)