Hi All, Thanks for the insight into current Drill usage.
Just to clarify one point, the discussion around UNION, LIST and REPEATED LIST was specifically for handling non-projected columns in EVF. Some history. Originally, the readers would either a) read all columns into vectors, then rely on a PROJECT operator to remove them, or b) use super-complex, reader-specific code to do the projection at read time (e.g. Parquet). To provide better performance, and reduce direct memory fragmentation, EVF handles the projection at read time. Readers read all columns and provide data to EVF, EVF ignores the unprojected columns. This is super easy for scalar columns. Moderately complex for arrays, maps and maps of arrays. It got quite messy for UNION, LIST and REPEATED LIST due to the complex, type-based, ever-changing nature of those structures. The thing that makes UNION, LIST and REPEATED LIST "special" is that they all are based on a dynamic type system: data can change types, either per row (UNION), a list of varying-type values (LIST), or lists of lists of varying-type values (REPEATED LIST). Following Visual Basic, I call these "variants": we have variant (UNION), variant arrays (LIST), and 2D variant arrays (REPEATED LIST). Oddly, there is no "REPEATED UNION": one uses LIST instead -- the only type in Drill with this distinction. We end up with REPEATED LIST only because we have REPEATED everything (except UNION), even though LIST itself is already a "REPEATED UNION." Yes, it is a bit of a mess and has been so from the start. The variant types, of course, were meant to allow reading unstructured JSON or XML: Drill could query a web page, say. But, of course, SQL is an awful language to work with such data. As a result, Drill's variant functionality has always been a muddle: it mostly works at the vector level (after numerous fixes), but SQL is not up to the task of wrangling the resulting data. To be fair, the Druid folks confidently ran into the same mess when they started evolving Druid to handle unstructured JSON data using SQL: it *should* work in theory, but doesn't actually work in practice. They ran into the same problems we've discussed: this suggests that the problems are real, not just the result of lack of knowledge by folks who built the thing. So, a proposal is to handle projection at read time for all but the variant cases. EVF always writes variant vectors, projected or not. A PROJECT operator removes unwanted columns later in the DAG, just like in the pre-EVF days. The result is not performant, but who cares if few people use these features. For the remaining readers not yet using EVF, I might suggest that if it ain't broke, don't fix it. The key value that EVF adds for existing readers is efficient handling of non-projected columns. However, for something like the JSON, Druid or Mongo readers, the DB itself will return only the projected columns, so the reader should project all returned columns. (Contrast this with a JSON file: we are forced to read all the columns, even if we only want 5 out of 100. Dumping the other 95 a read time is a big win.) Existing readers, such as Parquet, already contain one-off versions of the complex code needed to write to vectors and project columns. There is no win for the users by taking on the daunting challenge of rewriting the Parquet reader to use EVF. For NEW readers, of course, EVF provides all this functionality ready-made, so new readers should be written to use EVF. The key purpose of Parquet back in the halcyon days of Hadoop was as a compact, efficient data file format for distributed processing, especially for applications with thousands or millions of files. Impala, for example, targeted this exact use case. Does Parquet still have this role on large, cloud-based systems? If so, use Athena or Presto, along with Glue for metadata. For the desktop, use Drill without metadata. Still, is Parquet-at-scale a common format for desktop usage? If not, then the existing reader is probably good enough. Data "lakehouses" seem to be the new thing: Delta Lake with Spark and the like. Drill would need some significant work to play well in that context since it was designed for stand-alone usage and Hadoop integration. As Hadoop continues to fade, and new architectures emerge, perhaps Drill will avoid lakehouse integration work by focusing on the desktop/small-deployment use case. To summarize, for the use cases where Drill still finds users, we can punt on the variant types (UNION, LIST, REPEATED LIST): EVF can read them, but will always project them. A PROJECT operator can remove them if they are not projected. This approach has a performance hit, but that won't matter if no one actually uses these particular types. If we go this route, then the fixes I did are probably good enough: I just need to dust off the code and create a PR. Does this make sense? Thanks, - Paul On Mon, Jan 1, 2024 at 7:17 AM Charles Givre <cgi...@gmail.com> wrote: > To continue the thread hijacking.... > > I'd agree with what James is saying. What if we were to create a docker > container (or some sort of package) that included Drill, Superset and all > associated configuration stuff so that a user could just run a docker > command and have a fully functional Drill instance set up with Superset? > > Regarding the JSON, for a while we were working on updating all the > plugins to use EVF2. From my recollection, we got all the formats > converted except for parquet (major project) and HDF5 (PR pending: > https://github.com/apache/drill/pull/2515). We had also started working > on removing the old JSON reader, however, there were a few places it reared > its head: > 1. The Druid plugin. I wrote a draft PR that is pending to swap it out > for the EVF JSON reader but haven't touched it in a really long time. ( > https://github.com/apache/drill/pull/2657) > 2. The Mongo plugin: No work there... > 3. The conversion UDFs. Work started. ( > https://github.com/apache/drill/pull/2567) > > In any event, given the interest in Mongo/Drill, it might be worthwhile to > take a look at the Mongo plugin to see what it would take to swap out the > old JSON reader for the EVF one. > Regarding unprojected columns, if that's the holdup, I'd say scrap that > feature for complex data types. > > What do you think? > > > > On Jan 1, 2024, at 07:57, James Turton <dz...@apache.org> wrote: > > > > P.P.S. since I'm spamming this thread today. With > > > > > this suggests to me that we should keep putting effort into: embedded > Drill, Windows support, rapid installation and setup, low "time to insight". > > > > I'm not going so far as to suggest that Drill be thought of as desktop > software, rather that ad hoc Drill deployments working on small (Gb) to big > (Tb) data may be as, or more, important than long lived, heavily > integrated, professionally managed deployments working on really Big data > (Pb). Perhaps the last category belongs almost entirely to BigQuery, > Athena, Snowflake and the like nowadays anyway. > > > > I still think a cluster is the often the most effective way to deploy > Drill so the question contemplated is really "Can we make it faster and > easier to spin up a cluster (and embedded Drill), connect to data sources > and start running (successful) queries"? > > > > On 2024/01/01 07:33, James Turton wrote: > >> P.S. I also have an admittedly vague idea about deprecating the UNION > data type, which still breaks things in many operators, in favour of a > different approach where we kick any invalid data encountered while loading > column FOO out to a generated _FOO_EXCEPTIONS VARCHAR (or VARBINARY, though > binary data formats tend not to be malformed?) column. This would let a > query over dirty data complete without invisible data swallowing, and would > mean we could cut further effort on UNION support. > >> > >> On 2024/01/01 07:11, James Turton wrote: > >>> Happy New Year! > >>> > >>> Here's another two cents. Make that five now that I scan this email > again! > >>> > >>> Excluding our Docker Hub images (which are popular), Drill is > downloaded ~1000 times a month [1] (order of magnitude, it's hard to count > genuinely new installations from web server downloads). > >>> > >>> What roles are these folks in? I'm a data engineer by day and I don't > think that we count for a large share of those downloads. The DEs I work > with are risk averse sorts that tend to favour setups with rigid schemas > early on and no surprises for their users at query time. Add to that a > second stat from the download data: the biggest single download user OS is > Windows, at about 50% [1]. Some of these users may go on to copy that > download to a server environment but I have a theory that many of them go > on to run embedded Drill right there on beefy Windows laptops. > >>> > >>> I conjecture that most of the people reaching for Drill are analysts > or developers working _away_ from an established, shared data > infrastructure. There may not be any shared data engineering where they > are, or they may find themselves in a fashionable "Data Mesh" environment > [2]. I'm probably abusing Data Mesh a bit here in that I'm told that it > mainly proposes a federation of distinct data _teams_, rather than of data > _systems_ but, if you entertain my cynical formulation of "Data Mesh guys! > Silos aren't uncool any more!" just a bit, then you can well imagine why a > user in a Data Mesh might look for something like Drill to combine data > from different silos on their own machine. Tangentially this suggests to me > that we should keep putting effort into: embedded Drill, Windows support, > rapid installation and setup, low "time to insight". > >>> > >>> MongoDB questions still come up frequently giving a reason beyond the > JSON files questions to think that the JSON data model is still very > important. Wherever we decide to bound the current EVF v2 data model > implementation, maybe we can sketch out a design of whatever is > unimplemented in some updates to the Drill wiki pages? This would give > other devs a head start if we decide that some unsupported complex data > type is worth implementing down the road? > >>> > >>> 1. https://infra-reports.apache.org/#downloads&project=drill > >>> 2. https://martinfowler.com/articles/data-mesh-principles.html > >>> > >>> Regards > >>> James > >>> > >>> On 2024/01/01 03:16, Charles Givre wrote: > >>>> I'll throw my .02 here... As a user of Drill, I've only had the > occasion to use the Union once. However, when I used it, it consumed so > much memory, we ended up finding a workaround anyway and stopped using it. > Honestly, since we improved the implicit casting rules, I think Drill is a > lot smarter about how it reads data anyway. Bottom line, I do think we > could drop the union and repeated union. > >>>> > >>>> The repeated lists and maps however are unfortunately something that > does come up a bit. Honestly, I'm not sure what work is remaining here > but TBH Drill works pretty well at the moment with most of the data I'm > using it for. This would include some really nasty nested JSON objects. > >>>> > >>>> -- C > >>>> > >>>> > >>>>> On Dec 31, 2023, at 01:38, Paul Rogers <par0...@gmail.com> wrote: > >>>>> > >>>>> Hi Luoc, > >>>>> > >>>>> Thanks for reminding me about the EVF V2 work. I got mostly done > adding > >>>>> projection for complex types, then got busy on other projects. I've > yet to > >>>>> tackle the hard cases: unions, repeated unions and repeated lists > (which > >>>>> are, in fact, repeated repeated unions). > >>>>> > >>>>> The code to handle unprojected fields in these areas is getting > awfully > >>>>> complicated. In doing that work, and then seeing a trick that Druid > uses, > >>>>> I'm tempted to rework the projection bits of the code to use a > cleaner > >>>>> approach. However, it might be better to commit the work done thus > far so > >>>>> folks can use it before I wander off to take another approach. > >>>>> > >>>>> Then, I wondered if anyone actually still uses this stuff. Do you > still > >>>>> need the code to handle non-projection of complex types? > >>>>> > >>>>> Of course, perhaps no one will ever need the hard cases: I've never > been > >>>>> convinced that unions, repeated lists, or arrays of repeated lists > are > >>>>> things that any sane data engineer will want to use -- or use more > than > >>>>> once. > >>>>> > >>>>> Thanks, > >>>>> > >>>>> - Paul > >>>>> > >>>>> > >>>>> On Sat, Dec 30, 2023 at 10:26 PM James Turton <dz...@apache.org> > wrote: > >>>>> > >>>>>> Hi Luoc and Drill devs! > >>>>>> > >>>>>> It's best to email Paul directly since he doesn't follow these lists > >>>>>> closely. In the meantime I've prepared a PR of backported fixes for > >>>>>> 1.21.2 to the 1.21 branch [1]. I think we can try to get the Netty > >>>>>> upgrade that Maksym is working on, and which looks close to done, > >>>>>> included? There's at least one CVE applicable to our current > version of > >>>>>> Netty... > >>>>>> > >>>>>> Regards > >>>>>> James > >>>>>> > >>>>>> > >>>>>> 1. https://github.com/apache/drill/pull/2860 > >>>>>> > >>>>>> On 2023/12/11 04:41, luoc wrote: > >>>>>>> Hello all, > >>>>>>> 1.22 will be a more stable version. This is a digression: Is > Paul > >>>>>> still interested in participating in the EVF V2 refactoring in the > >>>>>> framework? I would like to offer time to assist him. > >>>>>>> luoc > >>>>>>> > >>>>>>>> 2023年12月9日 01:01，Charles Givre <cgi...@gmail.com> 写道： > >>>>>>>> > >>>>>>>> Hello all, > >>>>>>>> Happy Friday everyone! I wanted to raise the topic of getting a > Drill > >>>>>> minor release out the door before the end of the year. My opinion > is that > >>>>>> I'd really like to release Drill 1.22 once the integration with > Apache > >>>>>> Daffodil is complete, but it sounds like that is still a few weeks > away. > >>>>>>>> What does everyone think about issuing a maintenance release > before the > >>>>>> end of the year? There are a number of singificant fixes including > some > >>>>>> security updates and a major bug in the ES plugin that basically > makes it > >>>>>> unusable. > >>>>>>>> Best, > >>>>>>>> -- C > >>>>>> > >>> > >> > > > >
