- please see inline
SSSD is being very actively developed, adding several major features in each release. We need to make sure the code stays maintainable and adding new features in the upcoming release won’t increase the cost of maintaining SSSD long-term.
Since SSSD releases are primarily being driven by Fedora and RHEL releases, the Red Hat employed developers have a fixed amount of time for code refactoring. Of course, community members and developers are free to submit their patches on their schedule – although discussion on the list would be needed prior to merging any refactoring to not disrupt SSSD release quality for everyone.
A typical use-case would be: “A feature X depends on module Y that either is missing some functionality that is missing or a module that has outlived its initial design. Changing Y in that module would allow us to implement X more easily or with less maintenance effort in the future”.
The goal is to prepare the code for upcoming features without regressing, so testing after the refactoring is done is mandatory. We should consider also doing an upstream (pre)release to make it easier to test the changes.
This section lists the proposed tickets along with justifications, scope and test impact.
Given the fixed amount of time, each refactoring has a scope, expressed in just three high-level buckets - large (a couple of weeks, might take most of the time of the refactoring “sprint”), medium (a week to two weeks) or small (a couple of days, up to a week). Each item also lists the affected modules or functionality, so that we know where we need to improve tests.
Currently each responder (except the InfoPipe responder and several parts of the NSS responder) copy the logic that checks the cache and contacts the Data Provider if needed. In 1.15, we should add the missing functionality into the cache_req module and convert the existing responders (especially those that look up users and groups, not necessarily other objects like autofs maps or hosts..) to cache_req.
In 1.15, we should look at allowing lookups from trusted domain with a shortname. But we need to take performance into account and avoid cycling over all domains including their LDAP server. Then we could switch to checking the caches of all domains first before checking each domain’s cache and then its server.
This goal is tracked by <https://github.com/SSSD/sssd/issues/1885(Login time increases strongly if more than one domain is configured) and ultimately by <https://github.com/SSSD/sssd/issues/4042([RFE] Short name input format with SSSD for users from all domains when domain autodiscovery is used).
- cache_req: complete the needs of NSS responders
- Reuse cache_req() in responder code
We already have NSS and PAM responder tests. We need to extend them further to make sure all codepaths we change are tested.
sdap_async_groups.c module grew organically over time. At the moment, the module is quite hard to read and repeats some potentially expensive operations (like looping over all attributes or all members) several times.
In order to improve performance, we should refactor this module and test it extensively.
sdap_async_groups.c module would be better maintainable and we would remove some performance bottlenecks from the code.
- Refactor the sdap_async_groups.c module
LDAP group lookups can be tested using integration tests, “just” all cases we change must have corresponding test cases.
sdap_id_ops.c module was written in time where SSSD only supported a single domain. One of the things that are repeatedly biting us is that the module can set the fail over status of the whole domain to offline. Moreover, the module has no tests and is not easy to read.
At this time, it’s not clear whether the refactoring would just result in documenting and testing the module or if it would be worth for example making the module return error codes for connection errors and let the caller handle the errors. Alternatively, we might decide to do even more work and let the fail over code work per-domain, not per-back end, which probably wouldn’t be doable in the given scope. More research is needed.
The module would be better maintainable (currently there are some codepaths where we even don’t know why they were added anymore..), have tests and we would work towards removing issues with trusted domains setting SSSD to the offline mode.
- https://github.com/SSSD/sssd/issues/2549 - Investigate terminating connections in sdap_ops.c and comment the code some more
Other related tickets include:
- The sdap_op code always ends request with EAGAIN
- sdap code can mark the whole sssd_be offline
Currently upstream has only basic tests with the integration tests. Downstream has tests for fail over as well.
Medium to large, depending on what changes we decide to do.
As long as a request is confined within a single domain, we can pass around
sysdb_attrs or a similar data structure between different requests and avoid a costly cache writes. However, when a request must include processing in two different domain types, for example an IPA domain that includes overrides, the only way to pass intermediate data is to call a sysdb transaction and save the data to cache so that another request can read them.
Performance benefit in case SSSD must call identity lookup requests from different domains.
- https://github.com/SSSD/sssd/issues/3984 - Split out the requests for IPA users and groups that include overrides into reusable requests
Unfortunately, there are no upstream tests for requests that include overrides. Testing would be provided by downstream tests.
Medium to large
At the moment, we don’t have any upstream tests for the AD provider and we rely on downstream and manual testing completely. Nikolai Kondrashov wrote a proof-of-concept patches that provisions an AD DC server provided by the Samba project using the cwrap wrapper libraries. The scope of this effort would be to upstream this work.
SSSD integration tests would allow us to write tests for the AD provider.
- Investigate if Samba4 in Fedora can be used for SSSD CI
Some basic tests like looking up a user or a group would be part of this effort.
SSSD is gradually moving to socket-activated services and in general more self-contained services rather than implementing a service manager in the monitor process. The scope of this work would be to further decrease the dependence of services on the monitor process, such as moving the register functionality to the services themselves. Ultimately, the monitor process would perform one-time tasks such as converting the configuration from INI to confdb and exit.
Other work might include a fallback configuration or starting the services and domains even without having them explicitly enumerated in the services section.
Socket-activatable services would be better manageable by SSSD.
- RFE: Drop the monitor process
There are no upstream test for this functionality at the moment. Some service restart tests exist in downstream, though.
Medium to large, but hopefully this task could be split into several smaller tasks.
There are several improvements to the memory cache that we have been discussing lately, including a memory cache for by-SID lookups or having the memory cache respect case insensitive domains. The goal of this task would be to investigate what needs to be changed in the memory cache in order to implement these improvements.
Better performance through leveraging memory cache for SID lookups and lookups in case insensitive domains.
- [RFE] Support aliases in the memory cache
- Add a memcache for SID-by-id lookups
We already have tests for memory cache which could be extended. Tests for by-SID lookups would probably require us to add the Samba-based tests first.
Probably large, but more investigation is needed.
Our internal d-bus interface got a lot of new functionality to properly support D-Bus on public level. The InfoPipe responder has grown and also our internal communication between responders and providers has become more advanced.
The more we use it, the more it seems that the API that takes care of managing/terminating/error sbus requests is not optimal, since it requires a lot of glue code and often requires several output places and return code.
We should base sbus handlers on tevent to make sure there is only one output place and return code (when tevent request finishes) and we should also improve and simplify API that is used by handlers.
SSSD depends on D-Bus (and thus on sbus) more and more and we will keep adding new functionality. Reducing the amount of code that needs to be added and simplified logic will helps us to develop more stable code more quickly.
- none currently
Sbus is currently heavily tested. We may want to add new tests for new/changed API.
Failover mechanism wasn’t prepared for subdomains and we run into troubles every now and then. We added several workarounds for cases where the original code wasn’t sufficient but it made the code just more confused. At this moment nobody understands it but bugs keeps coming.
We should have a separate failover context for each domain, instead of one per whole backend. It must be possible to set different srv mechanism for each context. DNS resolver and cache should be shared between contexts.
We remove old and problematic code that nobody understands. We can improve site discovery for trusted domains and we can have better control over subdomain server resolution.
Downstream tests should remain functional, but upstream test will become invalid.
Probably out of four week scope.
Run all available upstream and downstream tests, if possible, extend the upstream tests.