Thanks everyone for the feedback. Looks like we are in general agreement. I am inclined to just do 1 & 2 and leave COPY_ON_WRITE as is based on great points Ethan and Shiyan raised. Makes sense.. Will wait for 1-2 days still to close this thread.
@semanticbeeing Thats a great idea. Is it more like a technical glossary of sorts? Lets may be start a different DISCUSS thread on that specific topic, so everyone can chime in and provide more attention to that proposal? On Tue, Nov 12, 2019 at 2:44 PM Y. Ethan Guo <[email protected]> wrote: > +1 on [1] and [2]. > > For [3], I have similar doubts as Shiyan. > > For the naming, I can understand the original intent of the analogy for COW > which is to make another "copy" of columnar/parquet file upon the > modification/update to the records in the file. From the system design > point of view, it's easy to understand. I'm okay with the renaming as > "MERGE_ON_WRITE" since it's probably straightforward for users at the first > glance. > > In terms of the concept, COW and MOR are listed as storage/table types. > From my understanding, they represent different tradeoffs of the > performance between reading and writing Hudi tables, and within MOR there > are different tradeoffs, e.g., lazy merge on read or periodic compaction > and cleaning pipelined along ingestion. It looks like these can be > controlled through configs, e.g., "disable_merge_on_write", > "compaction_frenquency", etc., instead of fixing the storage type, to > control the tradeoff that a user would like to make. The requirement may > change so a user can switch between COW and MOR by tuning the configs. We > don't have to make such changes now, but I'm wondering if this is something > worth considering in the future releases. > > - Ethan > > On Tue, Nov 12, 2019 at 8:43 AM nishith agarwal <[email protected]> > wrote: > > > +1 on the first two, don't feel strongly about (3). > > > > Thanks, > > Nishith > > > > On Tue, Nov 12, 2019 at 5:03 AM leesf <[email protected]> wrote: > > > > > [1] +1. `views` indeed confused me a lot. > > > [2] +1. `snapshot` is more reasonable. > > > [3] I don't feel very strong to rename it, the current name > > `COPY_ON_WRITE` > > > is reasonable considering the cost to rename and the behavior that new > > > version parquet file will be created and seems to be copied from old > > > version parquet file. > > > > > > Best, > > > Leesf > > > > > > Balaji Varadarajan <[email protected]> 于2019年11月12日周二 下午3:55写道： > > > > > > > Agree with all 3 changes. The naming now looks more consistent than > > > > earlier. +1 on them > > > > > > > > Depending on whether we are renaming Input formats for (1) and (2) - > > this > > > > could require some migration steps for > > > > > > > > Balaji.V > > > > > > > > > > > > On Mon, Nov 11, 2019 at 7:38 PM vino yang <[email protected]> > > wrote: > > > > > > > > > Hi Vinoth, > > > > > > > > > > Thanks for bringing these proposals. > > > > > > > > > > +1 on all three. Especially, big +1 on the third renaming proposal. > > > > > > > > > > When I was a newbie. The "COPY_ON_WRITE" term confused me a lot. It > > > > easily > > > > > mislead users on the "copy" term. And make users compare it with > the > > > > > `CopyOnWriteArrayList` data structure provided by JDK and thoughts > > of > > > > the > > > > > file systems. > > > > > > > > > > Best, > > > > > Vino > > > > > > > > > > > > > > > Bhavani Sudha <[email protected]> 于2019年11月12日周二 上午9:05写道： > > > > > > > > > > > +1 on all three rename proposals. I think this would make the > > > concepts > > > > > > super easy to follow for new users. > > > > > > > > > > > > If changing [3] seems to be a stretch, we should definitely do > [1] > > & > > > > [2] > > > > > at > > > > > > the least IMO. I will be glad to help out on the renames to > > whatever > > > > > extent > > > > > > possible should the Hudi community incline to pursue this. > > > > > > > > > > > > Thanks, > > > > > > Sudha > > > > > > > > > > > > > > > > > > > > > > > > On Mon, Nov 11, 2019 at 3:46 PM Vinoth Chandar < > [email protected]> > > > > > wrote: > > > > > > > > > > > > > Hello all, > > > > > > > > > > > > > > I wanted to raise an important topic with the community around > > > > whether > > > > > we > > > > > > > should rename some of our terminologies in code/docs to be more > > > > > > > user-friendly and understandable.. > > > > > > > > > > > > > > Let me also provide some context for each, since I am probably > > > guilty > > > > > of > > > > > > > introducing most of them in the first place :). > > > > > > > > > > > > > > *1. Rename "views" to "query" : *Instead of saying incremental > > view > > > > or > > > > > > > read-optimized view, talk about them as "incremental query" and > > > > > > > "read-optimized query". The term "view" is very technical, and > > > what I > > > > > was > > > > > > > trying to convey was that we ingest/store the data once and > > expose > > > > > views > > > > > > on > > > > > > > top. But new users (atleast half dozen of them to me) tend to > > > confuse > > > > > > this > > > > > > > with views/materialized views found in databases. Almost always > > we > > > > talk > > > > > > > about views mostly in terms of expected behavior for a query on > > the > > > > > > view. I > > > > > > > am proposing to just call these different query types since > its a > > > > more > > > > > > > universally accepted terminology and IMO clearer. > > > > > > > > > > > > > > *2. Rename "Read-Optimized/Realtime" views to Snapshot views + > > Have > > > > > > > Read-Optimized view only for MOR storage :* This one is > probably > > > the > > > > > > > trickiest. Hudi was always designed with MOR in mind, even as > we > > > were > > > > > > > working on COW storage and consequently we named the pure > parquet > > > > > backed > > > > > > > view as Read-Optimized, hoping to name parquet + avro based > view > > as > > > > > > > Write-Optimized. However, we opted to name it Realtime to > > emphasize > > > > the > > > > > > > data freshness aspect. In retrospect, the views should have not > > > been > > > > > > named > > > > > > > after their performance characteristics but rather the classes > of > > > > > queries > > > > > > > done on them and guarantees for those (point above #1). > Moreover, > > > > once > > > > > we > > > > > > > have parquet embedded into the log format, then the tradeoffs > may > > > not > > > > > be > > > > > > > the same anyways. > > > > > > > > > > > > > > So combining with the renaming proposed in #1, we would end up > > with > > > > the > > > > > > > following.. > > > > > > > > > > > > > > Copy-On-Write : > > > > > > > [Old] Read-Optimized View => [New] Snapshot Query > > > > > > > [Old] Incremental View => [New] Incremental Query > > > > > > > > > > > > > > Merge-On-Read: > > > > > > > [Old] Realtime View => [New] Snapshot Query > > > > > > > [Old] Incremental View => [New] Incremental Query > > > > > > > [Old] ReadOptimzied View => [New] Read-Optimized Query (since > it > > is > > > > > read > > > > > > > optimized compared to Snapshot query always, at the cost of > > staler > > > > > data) > > > > > > > > > > > > > > Both changes #1 & #2 could be simpler changes to just code > > > > references, > > > > > > docs > > > > > > > and configs.. we can support both string for sometime and > > deprecate > > > > > > > eventually since queries are stateless. > > > > > > > > > > > > > > *3. Rename COPY_ON_WRITE to MERGE_ON_WRITE :* Name originated > > since > > > > the > > > > > > > design was very similar to > > > > > > > https://urldefense.proofpoint.com/v2/url?u=https-3A__en.wikipedia.org_wiki_Copy-2Don-2Dwrite&d=DwIFaQ&c=r2dcLCtU9q6n0vrtnDw9vg&r=z456dQQXMUCz1m72nlkFQpylUpdOVMBG38x2peG1m44&m=m1yKGEwnAUe_FyIsWFAo-YVKyfq1nayItNGNc7iv8Yw&s=y9XF8-75xzGHY4yCbfVVWcIC1sbEXDxitqeAS2A6GoQ&e= > > > > > > > filesystems > > > > > > > & snapshotting and we once hoped to push some of this logic > into > > > the > > > > > > > storage itself, all in vain. but the name stuck, even though > once > > > we > > > > > had > > > > > > > MERGE_ON_READ the focus was often on merge costs etc, which the > > > name > > > > > > > COPY_ON_WRITE does not convey directly. I don't feel very > strong > > > > about > > > > > > this > > > > > > > and there is also cost to changing this since its persisted > > inside > > > > > > > > > > https://urldefense.proofpoint.com/v2/url?u=http-3A__hoodie.properties&d=DwIFaQ&c=r2dcLCtU9q6n0vrtnDw9vg&r=z456dQQXMUCz1m72nlkFQpylUpdOVMBG38x2peG1m44&m=m1yKGEwnAUe_FyIsWFAo-YVKyfq1nayItNGNc7iv8Yw&s=930ugGMXsrqzE-acg9nfeoePBmVjTRG3gD765ihEiqU&e= > > and we will support both strings internally in > > > code > > > > > for > > > > > > > backwards compatibility anyway > > > > > > > > > > > > > > Naming something is very hard (yes, try :)).I believe these > > changes > > > > > will > > > > > > > make the project simpler to understand for everyone out there. > We > > > > also > > > > > > have > > > > > > > tons of new people here, so I am also happy to let go, if its > > > already > > > > > > clear > > > > > > > :) > > > > > > > > > > > > > > Please use the bullet number when you share your feedback so we > > > know > > > > > what > > > > > > > the discussion is about. > > > > > > > > > > > > > > Thanks > > > > > > > Vinoth > > > > > > > > > > > > > > > > > > > > > > > > > > > >
