I'm just a lurker and aspiring contributor, but as a Spark user upgrading twice is very confusing and would cause many or most users to fail to upgrade successfully to Spark 4 on a first go. That seems like a very bad user experience. I thought it was worthwhile stating this out loud.
Russell On Wed, Mar 12, 2025 at 11:05 PM Xiao Li <gatorsm...@gmail.com> wrote: > this vote is to allow streaming queries which had been ever run in Spark >> 3.5.4 to be upgraded with Spark 4.0.x, "without having to be upgraded with >> Spark 3.5.5+ in prior". > > > In the history of Apache Spark, have we ever required users to upgrade to > the next maintenance release before moving to a new feature or major > release? > > Xiao > > Adam Binford <adam...@gmail.com> 于2025年3月11日周二 09:08写道： > >> +1 (non-binding) >> >> It's a pretty in the weeds issue with how Structured Streaming works >> under the hood that's kinda hard to understand if you're not familiar with >> it. The migration logic doesn't mean users can still use the old config, >> it's purely behind the scenes to fix checkpoint metadata in streams created >> in 3.5.4. The 5 lines of code it takes to address a weird edge case for >> certain users that's already gone from master shouldn't be a huge deal. >> >> On Tue, Mar 11, 2025 at 1:43 AM Yang Jie <yangji...@apache.org> wrote: >> >>> >>> To Sean, you're right, I'm very sorry. >>> >>> From the perspective of compatibility and migratability, I think we >>> should migrate this logic to 4.0.0 and keep it in the codebase for a longer >>> time (or permanently), because we can't predict which version users of >>> 3.5.4 will choose next. >>> >>> >>> I don't want to discuss the so-called vendor issue. >>> >>> I withdraw my previous -1. >>> >>> Jie Yang. >>> >>> On 2025/03/11 04:42:25 Wenchen Fan wrote: >>> > Guys, let’s be honest about what we’re discussing here. >>> > >>> > If this is a migration issue, why would we even need a vote? We’ve been >>> > consistently adding configurations to restore legacy behavior instead >>> of >>> > removing them because we understand the challenges of upgrading Spark >>> > versions. Our goal has always been to make upgrades easier, even if it >>> > means carrying some technical debt. I don’t think we want to change >>> that >>> > culture now. >>> > >>> > If the concern is about vendor names appearing in the codebase, then >>> why is >>> > it a big deal this time when vendor names are already present >>> elsewhere? If >>> > we’ve failed to follow a policy, let’s correct it, but can someone >>> point to >>> > the specific policy we’re violating? >>> > >>> > If the vote is about adding migration logic to ease the upgrade from >>> 3.5.4 >>> > to 4.0.0, then +1, why not? >>> > >>> > Thanks, >>> > Wenchen >>> > >>> > >>> > >>> > On Mon, Mar 10, 2025 at 8:49 PM Jungtaek Lim < >>> kabhwan.opensou...@gmail.com> >>> > wrote: >>> > >>> > > Well said, Sean. Sorry I made you keep around here since it might >>> not be >>> > > clearly stated. My bad. >>> > > >>> > > Yang, how could we ever tolerate the fact there are "other" >>> occurrences of >>> > > vendor names in the codebase? Please go and search "databricks" in >>> the >>> > > codebase and be surprised. >>> > > >>> > > If we believe that having vendor names in the codebase will increase >>> > > the occurrence of making mistakes, why didn't we have a discussion >>> thread >>> > > earlier to remove all occurrences altogether? This is super tricky >>> because >>> > > I can even start to argue we have "Apple" as a vendor name in Apache >>> Spark >>> > > codebase. I'm not saying we use "apple" in the test data. See >>> > > `isMacOnAppleSilicon` in Utils. Is it unavoidable? No, >>> `isMacOnMSeries` or >>> > > `isMacOnSilicon` is enough. >>> > > >>> > > We really need to draw a line where we disallow vendor names on it - >>> if >>> > > it's the entire codebase, I don't really think it is realistic. >>> > > >>> > > This was really a mistake, and it was definitely not from referring >>> to the >>> > > existing codebase. Not having a vendor name does not change anything >>> on the >>> > > chance of encountering this issue again. If we really care, we >>> should think >>> > > about style checking, which is the only viable way to catch the >>> mistake. >>> > > Again, I'd argue we have to have a bunch of vendor names in that >>> style >>> > > check, not just the problematic vendor name. >>> > > >>> > > >>> > > On Tue, Mar 11, 2025 at 12:17 PM Sean Owen <sro...@gmail.com> wrote: >>> > > >>> > >> Doesn't the migration code 'clear' the debt? >>> > >> The proposal is not to continue to support the config. >>> > >> I feel like people are not quite understanding the change, and >>> objecting >>> > >> to something that doesn't exist. >>> > >> It's a shame, as this seems like something not even worth >>> discussing. I >>> > >> don't know why this triggered this much discussion. We have kept >>> deprecated >>> > >> methods without blinking, which is in comparison much bigger. >>> > >> Can we maybe ask you review the actual change in question? >>> > >> >>> > >> On Mon, Mar 10, 2025, 10:02 PM Yang Jie <yangji...@apache.org> >>> wrote: >>> > >> >>> > >>> -1 >>> > >>> Remove migration logic of incorrect `spark.databricks.*` >>> configuration >>> > >>> in Spark 4.0.0 because I think this configuration was initially >>> introduced >>> > >>> accidentally in Spark 3.5.4, lacking a clear design intent. >>> Although the >>> > >>> immediate maintenance cost of retaining this configuration >>> currently seems >>> > >>> limited, as subsequent versions iterate and user habits form, it >>> may lead >>> > >>> to the continuous accumulation of technical debt. When users come >>> to view >>> > >>> this configuration as one that can be relied on long-term, future >>> removal >>> > >>> may face greater resistance from users and could potentially >>> become an >>> > >>> entrenched and redundant configuration in the codebase. Therefore, >>> promptly >>> > >>> correcting this historically accidental configuration not only >>> maintains >>> > >>> the normativity of the Spark configuration system but also prevents >>> > >>> unintended configurations from becoming de facto standards, thereby >>> > >>> reducing long-term maintenance risks. >>> > >>> >>> > >>> Jie Yang >>> > >>> >>> > >>> On 2025/03/10 14:52:52 Dongjoon Hyun wrote: >>> > >>> > -1 because there exists a feasible migration path for Apache >>> Spark >>> > >>> 3.5.4 via Apache Spark 3.5.5. >>> > >>> > >>> > >>> > It's obvious that this Databricks' mistake already causes a huge >>> > >>> communication cost in the Apache Spark community and is suggesting >>> a burden >>> > >>> to enforce us to handle at least two more PRs at 4.0.0 and 4.1.0. >>> > >>> > >>> > >>> > Given that, I don't think >>> > >>> > - This is an inevitable or >>> > >>> > - This is 0 cost >>> > >>> > >>> > >>> > Dongjoon. >>> > >>> > >>> > >>> > On 2025/03/10 12:46:16 Jungtaek Lim wrote: >>> > >>> > > Starting from my +1 (non-binding). >>> > >>> > > >>> > >>> > > In addition, I propose to retain migration logic till Spark >>> 4.1.x and >>> > >>> > > remove it in Spark 4.2.0. >>> > >>> > > >>> > >>> > > On Mon, Mar 10, 2025 at 9:44 PM Jungtaek Lim < >>> > >>> kabhwan.opensou...@gmail.com> >>> > >>> > > wrote: >>> > >>> > > >>> > >>> > > > Hi dev, >>> > >>> > > > >>> > >>> > > > Please vote to retain migration logic of incorrect >>> > >>> `spark.databricks.*` >>> > >>> > > > configuration in Spark 4.0.x. >>> > >>> > > > >>> > >>> > > > - DISCUSSION: >>> > >>> > > > >>> https://lists.apache.org/thread/xzk9729lsmo397crdtk14f74g8cyv4sr >>> > >>> > > > ([DISCUSS] Handling spark.databricks.* config being exposed >>> in >>> > >>> 3.5.4 in >>> > >>> > > > Spark 4.0.0+) >>> > >>> > > > >>> > >>> > > > Specifically, please review this post >>> > >>> > > > >>> https://lists.apache.org/thread/xtq1kjhsl4ohfon78z3wld2hmfm78t9k >>> > >>> which >>> > >>> > > > explains pros and cons about the proposal - proposal is about >>> > >>> "Option 1". >>> > >>> > > > >>> > >>> > > > Simply speaking, this vote is to allow streaming queries >>> which had >>> > >>> been >>> > >>> > > > ever run in Spark 3.5.4 to be upgraded with Spark 4.0.x, >>> "without >>> > >>> having to >>> > >>> > > > be upgraded with Spark 3.5.5+ in prior". If the vote passes, >>> we >>> > >>> will help >>> > >>> > > > users to have a smooth upgrade from Spark 3.5.4 to Spark >>> 4.0.x, >>> > >>> which would >>> > >>> > > > be almost 1 year. >>> > >>> > > > >>> > >>> > > > The (only) cons in this option is having to retain the >>> incorrect >>> > >>> > > > configuration name as "string" in the codebase a bit longer. >>> The >>> > >>> code >>> > >>> > > > complexity of migration logic is arguably trivial. (link >>> > >>> > > > < >>> > >>> >>> https://github.com/apache/spark/blob/4231d58245251a34ae80a38ea4bbf7d720caa439/sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/OffsetSeq.scala#L174-L183 >>> > >>> > >>> > >>> > > > ) >>> > >>> > > > >>> > >>> > > > This VOTE is for Spark 4.0.x, but if someone supports >>> including >>> > >>> migration >>> > >>> > > > logic to be longer than Spark 4.0.x, please cast +1 here and >>> leave >>> > >>> the >>> > >>> > > > desired last minor version of Spark to retain this migration >>> logic. >>> > >>> > > > >>> > >>> > > > The vote is open for the next 72 hours and passes if a >>> majority +1 >>> > >>> PMC >>> > >>> > > > votes are cast, with a minimum of 3 +1 votes. >>> > >>> > > > >>> > >>> > > > [ ] +1 Retain migration logic of incorrect >>> `spark.databricks.*` >>> > >>> > > > configuration in Spark 4.0.x >>> > >>> > > > [ ] -1 Remove migration logic of incorrect >>> `spark.databricks.*` >>> > >>> > > > configuration in Spark 4.0.0 because... >>> > >>> > > > >>> > >>> > > > Thanks! >>> > >>> > > > Jungtaek Lim (HeartSaVioR) >>> > >>> > > > >>> > >>> > > >>> > >>> > >>> > >>> > >>> --------------------------------------------------------------------- >>> > >>> > To unsubscribe e-mail: dev-unsubscr...@spark.apache.org >>> > >>> > >>> > >>> > >>> > >>> >>> > >>> >>> --------------------------------------------------------------------- >>> > >>> To unsubscribe e-mail: dev-unsubscr...@spark.apache.org >>> > >>> >>> > >>> >>> > >>> >>> --------------------------------------------------------------------- >>> To unsubscribe e-mail: dev-unsubscr...@spark.apache.org >>> >>> >> >> -- >> Adam Binford >> >
