Thanks Dmitri for kicking off this thread! I think even just laying out the design considerations in the form of a Q&A like you did here is great as a supplemental design artifact for posterity and this helps address the "documentation" questions I brought up in https://github.com/apache/polaris/pull/3268
Personally I'm okay with having it in the main shared admin tool, as long as we can do it in a way that avoids "monolithic code" scaling issues that can come up as the set of backend-specific things grows. I guess this is a good opportunity to start establishing the precedent for how to structure: 1. Hierarchical command syntax? Would it be like "java -jar admin-tool nosql maintenance garbage-collect --cel-expression='ageDays < 30'"? Or "java -jar admin-tool maintenance nosql garbage-collect" (maintenance before nosql, or is maintenance specific to nosql? would we collect common maintenance commands that are persistence-agnostic into the base maintenance subcommand?) 2. Should we have compile-time options that can choose which subfeatures to build in case there are issues with some subfeature that aren't applicable to the user? 3. Should we lay out the code for easy segregation as we scale? We may not want one directory that contains a SpannerMaintenance, an AliyunMaintenance, FoundationDbMaintenance, etc all next to each other I think layout aspects could probably be addressed in an incremental way though, so at least I don't have any hard stance on what's the right answer, as long as we're flexible in willingness to change the syntax to be more nested/organized when we see the need. For CEL, I do think it's trickier to evolve it to *take away* expressiveness in the future if we let the cat out of the bag to allow too expressive a language initially, since it pertains to the *semantics* of what people running the NoSQL impl come to depend on, beyond just *syntax* (i.e., it's somewhat easier to change the CLI's syntax to introduce a nesting like "admin-tool nosql maintenance garbage-collect" if the underlying functionality is the same, but if someone decides to start depending on being able to runtime-specify CEL expressions like 'getDayOfWeek(commitTime) == FRIDAY' it's hard to go back to a simpler world where we didn't have to deal with that). Note, I might be exaggerating my assumption about what the CEL expression supports here since I remember offhand the details about the part of the code that consumes it and I couldn't find docs on what we expect the structure of the input to the expression to be and what kind of CEL expressions are actually allowed. Your clarification that Polaris *Servers* won't need CEL on the classpath does help assuage my concerns about having it has a heavyweight dependency somewhat, but I think it's still prudent to know whether the intended use cases are a substantially more restrictive set of conditions (probably minimumNumToKeep and maxAge, right?). If we capture the pros/cons it'll help our future selves not have to redo the work in considering expressiveness vs precision/clarity of interface if someone tries to evolve the interface again in the future. I guess an argument in favor of CEL is that it's cumbersome/messier enough trying to express a combined numToKeep and maxAge condition in terms a multiple different config values that interact. On Thu, Jan 15, 2026 at 3:15 PM Dmitri Bourlatchkov <[email protected]> wrote: > Hi Prashant, > > You bring up valid points. Apologies for not explaining them upfront. I > suppose it's human nature to take things for granted when you've been > working with them for a while :) > > The need to retain more than just the latest state of the catalog is > primarily rooted in Disaster Recovery scenarios (specifically with NoSQL > persistence). > > In short, a DR situation may leave the latest state unusable (e.g. due to > replication lag... exact failures are kind of complex and probably require > a separate discussion), so the admin user may have to reset the catalog to > a previous state. This would be a data loss situation, of course, but it > may be the best option to recover some data compared to total loss. > > This is not actualized as specific user-level tools in OSS yet. Full DR > support requires considerable follow-up work. > > Whether the flexibility provided by CEL is really required for end users > can probably be debated. Let me think more about that. > > Re: CEL java maintainability, the Nessie CEL implements a particular > version of the CEL spec and passes Google's conformance tests. Therefore, > it is a correct CEL impl. Whether it needs to adopt newer spec revisions is > not really a maintenance burden in Polaris unless we want to always use the > latest CEL spec, which IMHO is not a requirement as the supported version > is already pretty expressive. Please consider that CEL is engaged only when > the user performs NoSQL maintenance, otherwise it is just a jar inside the > Admin Tool. Polaris Servers should not need CEL on the class path, AFAIK. > > Re: sync vs. async maintenance, sync cannot be reliable if you assume that > any node can be killed at any time (which is the reality in k8s). > > Re: exposing NoSQL-specific commands in the Admin Tool, I personally think > it is similar to supporting different storage technologies in the Catalog > config (e.g. GCS vs. S3). Polaris CLI has a multitude of options for the > union of them, but not all features of one storage type are applicable to > others. > > Cheers, > Dmitri. > > On Thu, Jan 15, 2026 at 1:08 PM Prashant Singh via dev < > [email protected]> wrote: > > > Thank you for starting the thread Dmitri ! > > Thank you Peirre for the response, I certainly missed this section of the > > design document. > > > > I believe I was expecting a design doc explaining why we want to > > selectively retain the entities which are not the current version as if > > NoSQL implementation cares about this, is there any design for this ? > > secondly as proposed in the doc we should just be cleaning all the > entities > > that are not current so I am unsure why we want to have age>=30 days kind > > of retention ? If we selectively want to > > retain, we need to have a design doc for it to explain use cases, agree > on > > user facing constructs and other, for example a possible interpretation > is > > can i go back to the state of the catalog as of 30 days ago ? > > I don't think Polaris supports undrop or time travel, and I don;t JDBC > will > > be able to support it, so I believe NoSQL's *default* behaviour should be > > delete everything that's not current. > > > > I can see the admin tool mentioned, but what I can't see in the > > presentation is this whole module, design trade off of sync vs async > > maintenance, user specific constructs, for example retention expression, > > why is it required. I believe > > those things warrant a design for themselves is my take. > > > > With that being said I totally understand NoSQL requires maintenance, > what > > I fail to understand is why does NoSQL require retention expressions ? > why > > can't everything that's not currently marked as a GC candidate, if the > > issue is we need this for > > debugging then we should just have a simple config saying keep the > latest X > > commits. To me it feels like we are opening for cases such as time travel > > and undrop without border agreement with the community. If we want to do > > these additional things and expose these extra constructs which > > I think are good to do, they can't be part of the polaris repo but would > be > > a good tool for polaris goodies. > > > > Hence was the request to open the discussion in the thread as well as > have > > a debate on where this tool would be, because Admin tool presently just > has > > bootstrap and purge which are supported by both the persistence but > > maintenance is just NoSQL specific > > and there is no way JDBC and IMHO it would be very confusing for end user > > to see i can't retain my catalog state as of 30 days in JDBC vs in NOSQL > so > > leaking this to admin tool, IMHO is not a good idea, but am open to > hearing > > others on why its is and how this concern is handled! > > > > Regarding the expression language introduction (I humbly disagree that we > > need one), I went till the 8th page of this projectnessie/cel-java [1] > this > > has just done dependency update where as googles/cel-java is something > > google developers are actively working and cel-java > > is an google's spec so i would rather use google/cel-java rather than > have > > a third party dependency of the same spec implementation which google > owns. > > > > With that being said I am open to hearing from others as to why such > > constructs should be present in the NoSQL specially retained staff age <= > > 30 ? > > > > On an orthogonal note : It would have been better if we would have had > > these discussions before we merged the PR. > > > > Thank you again Dmitri for starting this conversation, I really > appreciate > > it ! > > > > [1] > > https://github.com/apache/polaris/pull/3268#pullrequestreview-3576273215 > > > > Best, > > Prashant Singh > > > > On Thu, Jan 15, 2026 at 2:24 AM Pierre Laporte <[email protected]> > > wrote: > > > > > Hi Dmitri, thanks for the comprehensive recap. > > > > > > For "the newly added Maintenance module was not exposed in previous > docs > > > related to NoSQL", I wonder whether this is just a misunderstanding. > As > > > Prashant noted, in the NoSQL presentation that was run a couple of > times > > by > > > Adam [1], there is a mention of "A maintenance task in the Admin CLI > > > Tool". And the original design doc [2] also contains an explanation as > > to > > > why this is necessary for NoSQL in the "Handling no longer needed > > objects" > > > section. Am I missing something? > > > > > > Regarding the repository choice, I would like to emphasize the > potential > > > overhead in release management. Today, we have a manual release > process > > > that only spans the `apache/polaris` repository. And we have a > > > semi-automated release process that is tighly coupled with the > > > `apache/polaris` repository. Tightly coupled because it is implemented > > as > > > Github workflows within that repository. Let's consider the potential > > > impacts on release process and cadence. > > > > > > [1] > > > > > > > > > https://docs.google.com/presentation/d/1lX2EdvM0SeyuOdO_u1idlWfmnlH3hFE16JEyWo45Bdo/edit?slide=id.p24#slide=id.p24 > > > [2] > > > > > > > > > https://docs.google.com/document/d/1POUWe0xMZOBoaJ6Rgiw35ziEoc6OEYCiW7Zk6bR9H6M/edit?tab=t.0#heading=h.ccj3ewbhhhhy > > > -- > > > > > > Pierre > > > > > > > > > On Wed, Jan 14, 2026 at 11:18 PM Dmitri Bourlatchkov <[email protected] > > > > > wrote: > > > > > > > Hi All, > > > > > > > > As Prashant mentioned in GH [1], the newly added Maintenance module > was > > > not > > > > exposed in previous docs related to NoSQL. Let's use this email > thread > > to > > > > discuss it and possible concerns people may have. Below, I'm > providing > > > > rationale for topics, of which I am aware. Please feel free to start > > new > > > > threads dedicated to other concerns. Let's keep this discussion > focused > > > on > > > > the NoSQL maintenance functionality, though. > > > > > > > > * Why is this code necessary? > > > > > > > > NoSQL persistence is not transactional. Even normal commits leave > some > > > > amount of historical data in the database. Failed commits may leave > > > > remnants of preparatory data in the database too. > > > > > > > > If not cleaned up, this will lead to virtually indefinite growth of > > > > persisted data over time. > > > > > > > > Therefore, some periodic async cleanup is necessary. The maintenance > > code > > > > in PR [3268] provides fundamental code for performing this cleanup. > > > > > > > > * Why does it have to be in the main repo? > > > > > > > > The code in PR [3268] has to align tightly with the actual NoSQL > > > > Persistence implementation. It has to evolve in sync with the data > > model > > > of > > > > stored data. > > > > > > > > Therefore, it is logical to keep it in the same repo as the > mainstream > > > > NoSQL Persistence code. > > > > > > > > * Why is CEL required? > > > > > > > > CEL was chosen based on prior work when the NoSQL Persistence was > > > developed > > > > in private. It provides an efficient and expressive medium for admin > > > users > > > > to define NoSQL maintenance policies. > > > > > > > > * Why is the Nessie CEL java impl. used? > > > > > > > > The Nessie CEL java impl. predates the Google impl. and has been used > > in > > > > production for years under various projects (including Nessie > itself). > > > The > > > > developers of the NoSQL persistence are more certain of the runtime > > > > behavior of the Nessie CEL impl. than of Google's. Switching to > > Google's > > > > CEL java requires additional work. > > > > > > > > * Can we express maintenance policies in some other, non-CEL way? > > > > > > > > Generally yes. However, this requires extra work and analysis of UX > > > impact. > > > > If anyone has a concrete proposal for non-CEL maintenance policies, > > > ideas / > > > > PRs are welcome for discussion, of course. > > > > > > > > * Why does the Admin Tool has to have maintenance commands [3395]? > > > > > > > > This is to allow users of Apache Polaris binary distributions to > > perform > > > > maintenance should they choose NoSQL Persistence. The Admin Tool is a > > > > natural home for the maintenance CLI because it is in fact intended > to > > > > perform direct manipulation of the Polaris database, such as creating > > the > > > > schema and bootstrapping realms (existing functionality). > > > > > > > > * Can the maintenance command [3395] live in the polaris-tools repo? > > > > > > > > This would effectively require the Admin Tool to live in > polaris-tools, > > > > which seems to be against the recent move to unify Admin and Service > > > > binaries [3340]. > > > > > > > > * Can the maintenance code be invoked in some other way > > (non-Admin-CLI)? > > > > > > > > Yes. For example, it is possible to build docker images dedicated to > > > > running the maintenance tasks without using the Admin CLI. This is > not > > > > implemented in Apache Polaris yet. The Admin CLI appears to offer the > > > best > > > > UX for admin users with minimal developer effort. > > > > > > > > [1] > > > > > > https://github.com/apache/polaris/pull/3268#pullrequestreview-3576273215 > > > > > > > > [3340] https://github.com/apache/polaris/pull/3340 > > > > > > > > [3268] https://github.com/apache/polaris/pull/3268 > > > > > > > > [3395] https://github.com/apache/polaris/pull/3395 > > > > > > > > Thought? Comments? > > > > > > > > Cheers, > > > > Dmitri. > > > > > > > > > >
