diff options
author | Vratko Polak <vrpolak@cisco.com> | 2020-01-17 19:28:52 +0100 |
---|---|---|
committer | Vratko Polak <vrpolak@cisco.com> | 2020-04-16 09:26:26 +0000 |
commit | b8162d04aaebb5761d5c5cb1d7d3d53629aef126 (patch) | |
tree | 51e95986b0b2f68f7a777a38733d482a7a1f200f /docs | |
parent | daa48bd849e20b1452928b561398d23bad5bcc6b (diff) |
Flag day doc update
+ Include naming section.
+ Accomodate for the current usage:
+ Deactivating change can be uploaded late.
+ Nect activating change can be squashed there.
- Jira tickets and e-mail notifications
still listed as required.
+ Add links to real changes as examples.
+ Remove old text related to CRC comparison implementation
and some scenarois no longer supported.
+ No double spaces.
Change-Id: I33a574989fdfe81a2c53279fb9336165323a6219
Signed-off-by: Vratko Polak <vrpolak@cisco.com>
Diffstat (limited to 'docs')
-rw-r--r-- | docs/automating_vpp_api_flag_day.rst | 460 |
1 files changed, 241 insertions, 219 deletions
diff --git a/docs/automating_vpp_api_flag_day.rst b/docs/automating_vpp_api_flag_day.rst index b297e2ac3c..00cb6000d6 100644 --- a/docs/automating_vpp_api_flag_day.rst +++ b/docs/automating_vpp_api_flag_day.rst @@ -13,37 +13,37 @@ limitations under the License. -AUTOMATING VPP-API FLAG DAY IN CSIT -=================================== +VPP API Flag Day Algorithm +^^^^^^^^^^^^^^^^^^^^^^^^^^ -**ABSTRACT** +Abstract +~~~~~~~~ -This document is intended to provide a solution to the problem of +This document describes the current solution to the problem of automating the detection of VPP API changes which are not backwards -compatible with existing CSIT tests and to automate the "Flag Day" +compatible with existing CSIT tests, by defining the "Flag Day" process of deploying a new set of CSIT tests which are compatible with the new version of the VPP API without causing a halt to the -normal VPP/CSIT operational CI process. This shall initially be +normal VPP/CSIT operational CI process. This is initially limited to changes in \*.api files contained in the vpp repo. Eventually the detection algorithm could be extended to include other integration points such as "directory" structure of stats segment or PAPI python library dependencies. -**MOTIVATION** +Motivation +~~~~~~~~~~ Aside of per-release activities (release report), CSIT also provides testing that requires somewhat tight coupling to the latest (merged but not released) -VPP code. Currently, the coupling is not as tight as possible. -Instead of testing HEAD builds of VPP directly with HEAD CSIT code, -HEAD of one project is run with somewhat older codebase of the other project. -Definition of what is the older codebase to use is maintained by CSIT project. -For older CSIT codebase, the project creates so-called "oper" branches. -For older VPP codebase, CSIT master HEAD contains identifiers +VPP code. Currently, HEAD of one project is run against somewhat older codebase +of the other project. Definition of what is the older codebase to use +is maintained by CSIT project. For older CSIT codebase, there are so-called +"oper" branches. For older VPP codebase, CSIT master HEAD contains identifiers for "stable" VPP builds. Such older codebases are also used for verify jobs, where HEAD of the other project is replaced by the commit under review. One particular type of jobs useful for VPP development is trending jobs. -They test latests VPP build with oper branch of CSIT, +They test latests VPP build with latest oper branch of CSIT, and analytics is applied to detect regressions in preformance. For this to work properly, VPP project needs a warning against breaking the assumptions the current oper branch makes about VPP behavior. @@ -51,8 +51,8 @@ In the past, the most frequent type of such breakage was API change. Earlier attempts to create a process to minimize breakage have focused on creating a new verify job for VPP (called api-crc job) that -votes -1 on a change that affect CRC values for API messages CSIT uses. -The list of messages and CRC values (multiple values are allowed) +votes -1 on a change that affects CRC values for API messages CSIT uses. +The list of messages and CRC values (multiple "collections" are allowed) is maintained in CSIT repository (in oper branch). The process was less explicit on how should CSIT project maintain such list. As CSIT was not willing to support two incpompatible API messages @@ -66,62 +66,221 @@ because oper branch went out of sync with VPP HEAD codebase. This second type of failure prevents any merges to VPP for a long time (12 hours is the typical time, give time zone differences). -The current version of this document is still being somewhat vague -on the CSIT side of the process, but introduces a new requirement: -The api-crc job should not give false -1, under any reasonable circumstance. -That means, if a VPP change (nor any of its unmerged ancestor commits) -does not affect any CRC values for messages used by CSIT, --1 should only mean "rebase is needed", and rebasing to HEAD should result -in +1 from the api-crc job. - -**VPP API FLAG DAY ALGORITHM USE CASE** - -The following steps describe the use case of the proposed -implementation for automating the VPP API "Flag Day" algorithm: - -#. A VPP patch with VPP API changes is submitted to - gerrit for review. This usually means .api files are edited, - but a patch that affects the way CRCs are computed is also an API change. +The current version of this document introduces two new requirements. +Firstly, the api-crc job should not give false -1, under any +(reasonable) circumstances. That means, if a VPP change +(nor any of its unmerged ancestor commits) does not affect any CRC values +for messages used by CSIT, -1 should only mean "rebase is needed", +and rebasing to HEAD should result in +1 from the api-crc job. +Secondly, no more than one VPP change is allowed to be processed +(at the same time). + +Naming +~~~~~~ + +It is easier to define the process after chosing shorter names +for notions that need long definition. + +Note: Everytime a single job is mentioned, +in practice it can be a set of jobs covering parts of functionality. +A "run" of the set of jobs passes only if each job within the set +has been run (again) and passed. + +Jobs +---- + ++ A *vpp verify* job: Any job run automatically, and voting on open VPP changes. + Some verify jobs compile and package VPP for target operating system + and processor architecture, the packages are NOT archived (currently). + They should be cached somewhere in future to speed up in downstream jobs, + but currently each such downstream job can clone and build. + ++ The *api-crc* job: Quick verify job for VPP changes, that accesses + CSIT repository (checkout latest oper branch HEAD) to figure out + whether merging the change is safe from CSIT point of view. + Here, -1 means CSIT is not ready. +1 means CSIT looks to be ready + for the new CRC values, but there still may be failures on real tests. + ++ A *trending* job: Any job that is started by timer and performs testing. + It checkouts CSIT latest oper branch HEAD, downloads the most recent + completely uploaded VPP package, and unconditionally runs the tests. + CRC checks are optional, ideally only written to console log + without otherwise affecting the test cases. + ++ A *vpp-csit* job: A slower verify job for VPP changes, that accesses CSIT + repository and runs tests from the correct CSIT commit (chosen as in trending) + against the VPP (built from the VPP patch under review). + Vote -1 means there were test failures. +1 means no test failures, meaning + there either was no API change, or it was backward compatible. + ++ A *csit-vpp* job: Verify job for open CSIT changes. Downloads the + (completely uploaded) VPP package marked as "stable", and runs a selection + of tests (from the CSIT patch under review). + Vote +1 means all tests have passed, so it is safe to merge + the patch under review. + ++ A *patch-on-patch* job: Manually triggered non-voting job + for open CSIT changes. Compiles and packages from VPP source + (usually of an unmerged change). Then runs the same tests as csit-vpp job. + This job is used to prove the CSIT patch under review is supporting + the specified VPP code. + In practice, this can be a vpp-csit job started with CSIT_REF set. + ++ A *manual verification* is done by a CSIT committer, locally executing steps + equivalent to the patch-on-patch job. This can to save time and resources. + +CRC Collections +--------------- + +Any commit in/for the CSIT repository contains a file (supported_crcs.yaml), +which contains either one or two collections. A collection is a mapping +that maps API message name to its CRC value. + +A collection name specifies which VPP build is this collection for. +An API message name is present in a collection if and only if +it is used by a test implementation (can be in different CSIT commit) +targeted at the VPP build (pointed out by the collection name). + ++ The *stable collection*: Usually required, listed first, has comments and name + pointing to the VPP build this CSIT commit marks as stable. + The stable collection is only missing in deactivating changes (see below) + when not mergeable yet. + ++ The *active collection*: Optional, listed second, has comments and name + pointing to the VPP Gerrit (including patch set number) + the currently active API process is processing. + The patch set number part can be behind the actual Gerrit state. + This is safe, because api-crc job on the active API change will fail + if the older patch is no longer API-equivalent to the newer patch. + +Changes +------- + ++ An *API change*: The name for any Gerrit Change for VPP repository + that does not pass api-crc job right away, and needs this whole process. + This usually means .api files are edited, but a patch that affects + the way CRC values are computed is also an API change. + + Full name could be VPP API Change, but as no CSIT change is named "API change" + (and this document does not talk about other FD.io or external projects), + "API change" is shorter. + + TODO: Is there a magic incantation for Gerrit WebUI to search for API changes? + Open, -1 from api-crc job, +1 from other (non-csit) jobs. + ++ A *blocked change*: The name for open Gerrit Change for VPP repository + that got -1 from some of voting verify jobs. + ++ A *VPP-blocked change": A blocked change which got -1 from some "pure VPP" + verify job, meaning no CSIT code has been involved in the vote. + Example: "make test" fails. + + VPP contributor is expected to fix the change, or VPP developers + are expected to found a cause in an earlier VPP change, and fix it. + No interaction with CSIT developers is necessary. + ++ A *CSIT-blocked change*: A blocked change which is not VPP-blocked, + but does not pass some vpp-csit job. + To fix a CSIT-blocked change, an interaction with a CSIT committer + is usually necessary. Even if a VPP developer is experienced enough + to identify the cause of the failure, a merge to CSIT is usually needed + for a full fix. + + This process does not specify what to do with CSIT-blocked changes + that are not also API changes. + ++ The *active API change*: The API change currently being processed + by the API Flag Day Algorithm. + While many API changes can be open (waiting/queued/scheduled), + only one is allowed be active at a time. + ++ The *activating change*: The name for a Gerrit Change for CSIT repository + that does not change the test code, but adds the active CRC collection. + Merge of the opening change (to latest CSIT oper branch) defines + which API change has become active. + ++ The *deactivating change*: The name for Gerrit Change for CSIT repository + that only supports tests and CRC values for VPP with the active API change. + That implies the previously stable CRC collection is deleted, + and any edits to the test implementation are done here. + ++ The *mergeable deactivating change*: The deactivating change with additional + requirements. Details on the requirements are listed in the next section. + Merging this change finishes the process for the active API change. + +It is possible for a single CSIT change to act both as a mergeable +deactivating change for one API change, and as an activating change +for another API change. As English lacks a good adjective for such a thing, +this document does not name this change. +When this documents says a change is activating or deactivating, +it allows the possibility for the change to fullfill also other purposes +(e.g. acting as deactivating / activating change for another API change). + +Algorithm Steps +~~~~~~~~~~~~~~~ + +The following steps describe the application of the API "Flag Day" algorithm: + +#. A VPP patch for an API change is submitted to + gerrit for review. #. The api-crc job detects the API CRC values have changed for some messages used by CSIT. - Note: This requires a new CSIT test described below in the - section "CSIT VPP API CHANGE DETECTION TEST". -#. The api-crc job runs in parallel with any other VPP verify job, - so other jobs can hint at the impact on CSIT. - Some API changes do not need any edit on CSIT side, except for the CRC value. - If the api-crc job fails, an email with the appropriate reason +#. The api-crc job runs in parallel with any other vpp-csit verify job, + so those other jobs can hint at the impact on CSIT. + Currently, any such vpp-csit job is non-voting, + as the current process does not guarantee such jobs passes + when the API change is merged. +#. If the api-crc job fails, an email with the appropriate reason is sent to the VPP patch submitter and vpp-api-dev@lists.fd.io - including the VPP patch information and API change(s) that were detected. + including the VPP patch information and .api files that are edited. #. The VPP patch developer and CSIT team create a CSIT JIRA ticket to identify the work required to support the new VPP API version. -#. CSIT developer creates a patch to existing CSIT tests to support - the new VPP API version and new features in the VPP patch as required. -#. CSIT developer runs CI verification tests (without CRC checks) - for the CSIT patch against the VPP patch (aka "Patch-On-Patch verification"). - This process verifies support for the new VPP API, associated VPP - features and CSIT tests. Both developers iterate until the - verification passes. -#. CSIT developer creates a separate change, that only edits the CSIT CRC list - to include both stable VPP CRCs, and CRCs of the VPP change under review. -#. CSIT committer requires proofs both CRCs are included correctly - (for example the api-crc job is run with overriden CSIT_REF parameter). -#. When CSIT committer sees both CSIT changes are ready, - the CRC-editing change is merged to CSIT master branch - and cherry-picked to the latest oper branch. +#. CSIT developer creates a patch of the deactivating change + (upload to Gerrit not required yet). +#. CSIT developer runs patch-on-patch job (or manual verification). + Both developers iterate until the verification passes. + Note that in this phase csit-vpp job is expected to vote -1, + as the deactivating change is not mergeable yet. +#. CSIT developer creates the activating change, uploads to Gerrit, + waits for vote (usual review cycle applies). +#. When CSIT committer is satisfied, the activating change is merged + to CSIT master branch and cherry-picked to the latest oper branch. + This enters a "critical section" of the process. + Merges of other activating changes are not allowed from now on. + The targeted API change becomes the active API change. This does not break any jobs. -#. VPP developer issues a recheck on the VPP patch -#. On success, VPP Committer merges the patch. +#. VPP developer (or CSIT committer) issues a recheck on the VPP patch. +#. On failure, VPP and CSIT committers analyze what went wrong. + Typically, the active CRC collection is matching only an older patch set, + but a newer patch set needs different CRC values. + Either due to improvements on the VPP change in question, + or due to a rebase over previously merged (unrelated) API change. + VPP perhaps needs to rebase, and CSIT definitely needs + to merge edits to the active collection. Then issue a recheck again, + and iterate until success. +#. On success, VPP Committer merges the active API change patch. + (This is also a delayed verification of the current active CRC collection.) #. VPP committer sends an e-mail to vpp-api-dev stating the support for - the previous CRC values will soon be removed. + the previous CRC values will soon be removed, implying other changes + (whether API or not) should be rebased soon. #. VPP merge jobs create and upload new VPP packages. This breaks trending jobs, but both VPP and CSIT verify jobs still work. -#. CSIT developer bumps stable vpp version in the test affecting change - and rechecks it still works. -#. CSIT committer merges this change (to both master and oper). - This fixes trending jobs. Both VPP and CSIT verify jobs continue to work. -#. CSIT committer creates a patch that removes the old CRC values. -#. CSIT committer merges that patch. - This breaks verify jobs for old changes in VPP. +#. CSIT developer makes the deactivating change mergeable: + The stable VPP build indicator is bumped to the build + that contains the active API change. The active CRC collection + (added by the activating change) is renamed to the new stable collection. + (The previous stable collection has already been deleted.) + At this time, the deactivating change should be uploaded to Gerrit and + csit verify jobs should be triggered. +#. CSIT committer reviews the code, perhaps triggering any additional jobs + needed to verify the tests using the edited APIs are still working. +#. When satisfied, CSIT committer merges the mergeable deactivating change + (to both master and oper). + The merge fixes trending jobs. VPP and CSIT verify jobs continue to work. + The merge also breaks some verify jobs for old changes in VPP, + as announced when the active API change was merged. + The merge is the point where the process leaves the "critical section", + thus allowing merges of activating changes for other API changes. #. CSIT committer sends an e-mail to vpp-api-dev stating the support for the previous CRC values has been removed, and rebase is needed for all affected VPP changes. @@ -130,160 +289,23 @@ implementation for automating the VPP API "Flag Day" algorithm: submitter to rebase the patch to pick up the compatible VPP API version files. -**FEATURES REQUIRED FOR IMPLEMENTATION** +Real life examples +~~~~~~~~~~~~~~~~~~ + +Simple API change: https://gerrit.fd.io/r/c/vpp/+/23829 + +Activating change: https://gerrit.fd.io/r/c/csit/+/23956 + +Mergeable deactivating change: https://gerrit.fd.io/r/c/csit/+/24280 + +Less straightforward mergeable deactivating change: +https://gerrit.fd.io/r/c/csit/+/22526 +It shows: + ++ Crc edits: supported_crcs.yaml ++ Version bump: VPP_STABLE_VER_UBUNTU_BIONIC ++ And even a way to work around failing tests: + eth2p-ethicmpv4-ip4base-eth-1tap-dev.robot -**VPP API SIGNATURE GENERATION** - -The VPP PAPI generation already produces the complete set of -signatures in JSON format for all api files and includes them in the -vpp-api-python.deb package. Upon installation all of the \*.api.json -files are installed in the /usr/share/vpp/api directory. Each record -in the .api.json file contains the name of the api message, the fields -and their data types, and a CRC of the json object. - -**VPP API CLIENT SIGNATURES** - -In each CSIT branch, all of the VPP API client signatures that are supported -by the CSIT tests in that branch are contained in separate directories -under the .../csit/resources/api/vpp directory. The CSIT VPP API -client signature directory structure is the same as the one published in -/usr/share/vpp/api as generated by vppapigen. - -The granularity of the CSIT VPP API client signature support -will initially be on a per VPP API JSON file. In the future, a per VPP -api message level of granularity may be added. If CSIT is capable of -supporting more than one version of a VPP API JSON file, then a new -CSIT VPP api client signature directory will be created containing -all of the supported VPP API JSON files. Typically this will be identical -to the previous version with the exception of the VPP API JSON files -which have been changed in the VPP patch which triggered the VPP API FLAG -day algorithm. - -See https://gerrit.fd.io/r/19027 for the baseline implementation. - -**VPP API CHANGES FILE INCLUDED in VPP PACKAGE** - -The VPP build system shall add a file in the /usr/share/vpp/api -directory of the vpp package which is the same directory in which -the api JSON files are published. This file will include the list of -VPP api files which were included in the patch to be verified. - -See https://gerrit.fd.io/r/19479 for the baseline implementation. - -**CSIT VPP API CHANGE DETECTION TEST** - -The set of VPP api signatures which are supported by the CSIT tests in -a given CSIT branch shall be stored in .../csit/resources/api/vpp which -mirrors the same directory structure as the API signature directory -generated by vppapigen (e.g. /usr/share/api/vpp/core & -/usr/share/api/vpp/plugins). - -The test compares the VPP patch's API signature directory with each of -the CSIT VPP API signabture directory and determine the following state: - -- No Change -- Changed -- Rebase or Merge Parent VPP Patch [0] - -[0] The Rebase or Merge Parent VPP Patch result occurs when there is no valid API -signature found in .../csit/resources/api/vpp AND there are no VPP API changes -included in the patch. This could be the result of a patch whose parent does not -include the API changes merged in another VPP patch and supported by the new CSIT -oper branch. This case would be resolved by rebasing the patch to HEAD. The other -possibility is that the patch is a descendent of a patch with an incompatible API -change that has not been merged yet. This case is resolved by completing the API -Flag Day algorithm on the parent patch such that the latest CSIT oper branch supports -the API in the parent. This importance of the detection of this state is to provide -direct feedback to the VPP patch author about how to resolve the issue in a timely -manner. - -Any condition other than "No Change" shall cause an email to be sent -to the VPP patch submitter. If the condition is "Changed" then -vpp-api-dev@lists.fd.io shall also be copied on the notification email. - -**RUN CSIT VERIFY JOB AGAINST A SPECIFIC VPP PATCH IN GERRIT REVIEW BRANCH** - -This is the "Patch-On-Patch" methodology documented in [TBD]? - - -**VPP API FLAG DAY SCENARIOS** - -In the beginning, let's assume there is a single VPP API Client signature -directory in the current oper branch called vpp-api-client.sig.1 which -contains core/vpe.api.json and plugin/acl.api.json which are supported -by the CSIT tests. - -**VPP PATCH CONTAINS INCOMPATIBLE API CHANGES** - -Next, a VPP developer modifies vpe.api with a whole set of -new type definitions. When the patch is submitted to gerrit.fd.io, the -"CSIT VPP API CHANGE DETECTION TEST" detects the changed api file and -votes Verified -1. Once CSIT has been updated to support the new type -definitions and verified against the VPP patch, -vpp-api-client.sig.1/core/vpe.api.json is replaced with the vpe.api.json -file from the patch. The CSIT changes are committed into CSIT master and a -new oper branch is created. The VPP patch is then rechecked and merged -into VPP master as soon as practicable. All existing VPP patches and any -new patches not including the VPP api change patch will fail verification -with a "Rebase or Merge Parent" notification upon recheck or initial -submission to gerrit. Rebasing is then required in order to pass -verification of the new api changes. - -**VPP PATCH CONTAINS BACKWARDS COMPATIBLE CHANGES** - -The next day, a VPP developer finds a need to add a new -attribute to an api message in vpe.api with a default value defined. -This is a backwards compatible change for CSIT. Since the "CSIT VPP -API CHANGE DETECTION TEST" only works on a per api file level of granularity, -the change is flagged with Verified -1. However, in this case, the -CSIT developer can resolve the verify failure by adding a second VPP API -client signature directory, vpp-api-client.sig.2 which is a copy of -vpp-api-client.sig.1 with the vpe.api.json file updated with the contents -of the copy from the VPP patch. After the CSIT changes are merged and a new -CSIT oper branch is created, the VPP patch will pass verification upon recheck. -All other patches will continue to pass verification upon recheck or initial -submission to gerrit by matching the signature in vpp-api-client.sig.1 -- -life is good. - -**CSIT REMOVES SUPPORT FOR A VPP API VERSION** - -Since it is not desirable to maintain a bazillion CSIT VPP API client -signatures, after a reasonable period of time (let's say a week), a -CSIT developer deletes vpp-api-client.sig.1 and renames -vpp-api-client.sig.2 to vpp-api-client.sig.1, merges to CSIT master, -and creates a new oper branch. At this point, VPP patches that do not -contain the new vpe.api file will fail verification upon recheck or initial -submission to gerrit with a "Rebase or Merge Parent" notification and -will require rebasing to pass verification. - -**CSIT ADDS SUPPORT FOR A NEW FEATURE API PRIOR TO VPP** - -A VPP developer has lots of ideas and decides to add a new -plugin and api which supports the "Super-Duper Feature" to VPP in -a new plugin called the "Super-Duper Plugin" and associated super_duper.api -VPP binary APi message definition file. Being a thoughtful and -helpful developer, the VPP developer notifies the CSIT team providing -them with the super_duper.api.json file. A CSIT developer -quickly produces the Super-Duper Feature CSIT test suite and updates the VPP -API Client signature with vpe-api-client.sig.1/plugin/super_duper.api.json. -In the meantime, the VPP developer pushes the Super-Duper VPP patch which -fails the CSIT VPP API CHANGE DETECTION TEST. Both developers then work -together to verify both CSIT and the VPP patch. The CSIT developer -then merges the CSIT code into master and creates a new oper branch. Our -VPP developer is very pleased when the VPP patch containing -the Super-Duper Plugin verifies upon recheck. All other VPP patches without -api file changes continue to pass the CSIT VPP API CHANGE DETECTION TEST -before and after the Super-Duper VPP patch is merged. - -**VPP PATCH CONTAINS A NEW FEATURE API BEFORE CSIT SUPPORT** - -Now let's assume that the VPP developer was having a bad day -and forgot to notify the CSIT team about the new Super-Duper Plugin. -Upon pushing the VPP patch to gerrit, the VPP developer is pleased that -there is no nastygram email from the CSIT VPP API CHANGE DETECTION TEST. -All VPP patches without api file changes continue to pass the CSIT VPP -API CHANGE DETECTION TEST. Eventually a Super-Duper Plugin test suite is -added to CSIT along with vpe-api-client.sig.1/plugin/super_duper.api.json -and release in a new CSIT oper branch. All VPP patches that are do not contain -api changes and are verified via recheck or initial submission, continue to -pass the CSIT VPP API CHANGE DETECTION TEST. +Simple change that is both deactivating and activating: +https://gerrit.fd.io/r/c/csit/+/23969 |