thanks Adar for your comments and thinking this through! I really think Kudu has tons of potential and very happy we've made a decision to use it for our real-time pipeline.
you asked about use cases for support of nested and large binary/text objects. I work for a large healthcare organization and there is a lot of going with newer HL7 FHIR standard. FHIR documents are highly nested json objects that might get as big as a few Gbs in size. We could not use Kudu for storing FHIR bundles/documents due to the size and inability to store FHIR objects natively. We ended up using Hive/Spark for that, but that solution is not real-time. And it is not only about storing, but also about fast search in those FHIR documents. I know we could use Hbase for that but Hbase is really bad when it comes to analytics type of workloads. As for large text/binaries, the good and common example is narratives notes (physician notes, progress notes, etc.) They can be in form of RTF or PDF documents or images. 300 column limit was not a big deal so far but we have high-density nodes with 2 44-core cpus and 12 12-Tb drives and we cannot use the full power of them due to limitations of Kudu in terms of number of tablets per node. This is more concerning than 300 column limit. On Sat, Mar 14, 2020 at 3:45 AM Adar Lieber-Dembo <a...@cloudera.com> wrote: > Snowflake's micro-partitions sound an awful lot like Kudu's rowsets: > 1. Both are created transparently and automatically as data is inserted. > 2. Both may overlap with one another based on the sort column (Snowflake > lets you choose the sort column; in Kudu this is always the PK). > 3. Both are pruned at scan time if the scan's predicates allow it. > 4. In Kudu, a tablet with a large clustering depth (known as "average > rowset height") will be a likely compaction candidate. It's not clear to me > if Snowflake has a process to automatically compact their micro-partitions. > > I don't understand how this solves the partitioning problem though: the > doc doesn't explain how micro-partitions are mapped to nodes, or if/when > micro-partitions are load balanced across the cluster. I'd be curious to > learn more about this. > > Below you'll find some more info on the other issues you raised: > - Nested data type support is tracked in KUDU-1261. We've had an > on-again/off-again relationship with it: it's a significant amount of work > to implement, so we're hesitant to commit unless there's also significant > need. Many users have been successful in "shredding" their nested types > across many Kudu columns, which is one of the reasons we're actively > working on increasing the number of supported columns above 300. > - We're also working on auto-rebalancing right now; see KUDU-2780 for more > details. > - Large cell support is tracked in KUDU-2874, but there doesn't seem to > have been as much interest there. > - The compaction issue you're describing sounds like KUDU-1400, which was > indeed fixed in 1.9. This was our bad; for quite some time we were focused > on optimizing for heavy workloads and didn't pay attention to Kudu > performance in "trickling" scenarios, especially not over a long period of > time. That's also why we didn't think to advertise --flush_threshold_secs, > or even to change its default value (which is still 2 minutes: far too > short if we hadn't fixed KUDU-1400); we just weren't very aware of how > widespread of a problem it was. > > > On Fri, Mar 13, 2020 at 6:22 AM Boris Tyukin <bo...@boristyukin.com> > wrote: > >> thanks Adar, I am a big fan of Kudu and I see a lot of potential in Kudu, >> and I do not think snowflake deserves that much credit and noise. But they >> do have great ideas that take their engine to the next level. I do not know >> architecture-wise how it works but here is the best doc I found regarding >> micro partitions: >> >> https://docs.snowflake.net/manuals/user-guide/tables-clustering-micropartitions.html >> >> >> It is also appealing to me that they support nested data types with ease >> (and they are still very efficient to query/filter by), they scale workload >> easily (Kudu requires manual rebalancing and it is VERY slow as I've >> learned last weekend by doing it on our dev cluster). Also, they support >> very large blobs and text fields that Kudu does not. >> >> Actually, I have a much longer wish list on the Impala side than Kudu - >> it feels that Kudu itself is very close to Snowflake but lacks these >> self-management features. While Impala is certainly one of the fastest SQL >> engines I've seen, we struggle with mutli-tenancy and complex queries, that >> Hive would run like a champ. >> >> As for Kudu's compaction process - it was largely an issue with 1.5 and I >> believed was addressed in 1.9. We had a terrible time for a few weeks when >> frequent updates/delete almost froze all our queries to crawl. A lot of >> smaller tables had a huge number of tiny rowsets causing massive scans and >> freezing the entire Kudu cluster. We had a good discussion on slack and >> Todd Lipcon suggested a good workaround using flush_threshold_secs till we >> move to 1.9 and it worked fine. Nowhere in the documentation, it was >> suggested to set this flag and actually it was one of these "use it at your >> own risk" flags. >> >> >> >> On Thu, Mar 12, 2020 at 8:49 PM Adar Lieber-Dembo <a...@cloudera.com> >> wrote: >> >>> This has been an excellent discussion to follow, with very useful >>> feedback. Thank you for that. >>> >>> Boris, if I can try to summarize your position, it's that manual >>> partitioning doesn't scale when dealing with hundreds of (small) tables and >>> when you don't control the PK of each table. The Kudu schema design guide >>> would advise you to hash partition such tables, and, following Andrew's >>> recommendation, have at least one partition per tserver. Except that given >>> enough tables, you'll eventually overwhelm any cluster based on the sheer >>> number of partitions per tserver. And if you reduce the number of >>> partitions per table, you open yourself up to potential hotspotting if the >>> per-table load isn't even. Is that correct? >>> >>> I can think of several improvements we could make here: >>> 1. Support splitting/merging of range partitions, first manually and >>> later automatically based on load. This is tracked in KUDU-437 and >>> KUDU-441. Both are complicated, and since we've gotten a lot of mileage out >>> of static range/hash partitioning, there hasn't been much traction on those >>> JIRAs. >>> 2. Improve server efficiency when hosting large numbers of replicas, so >>> that you can blanket your cluster with maximally hash-partitioned tables. >>> We did some work here a couple years ago (see KUDU-1967) but there's >>> clearly much more that we can do. Of note, we previously discussed >>> implementing Multi-Raft ( >>> https://issues.apache.org/jira/browse/KUDU-1913?focusedCommentId=15967031&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-15967031 >>> ). >>> 3. Add support for consistent hashing. I'm not super familiar with the >>> details but YugabyteDB (based on Kudu) has seen success here. Their recent >>> blog post on sharding is a good (and related) read: >>> https://blog.yugabyte.com/four-data-sharding-strategies-we-analyzed-in-building-a-distributed-sql-database/ >>> >>> Since you're clearly more familiar with Snowflake than I, can you >>> comment more on how their partitioning system works? Sounds like it's quite >>> automated; is that ever problematic when it decides to repartition during a >>> workload? The YugabyteDB blog post talks about heuristics "reacting" to >>> hotspotting by load balancing, and concludes that hotspots can move around >>> too quickly for a load balancer to keep up. >>> >>> Separately, you mentioned having to manage Kudu's compaction process. >>> Could you go into more detail here? >>> >>> On Wed, Mar 11, 2020 at 6:49 AM Boris Tyukin <bo...@boristyukin.com> >>> wrote: >>> >>>> thanks Cliff, this is really good info. I am tempted to do the >>>> benchmarks myself but need to find a sponsor :) Snowflake gets A LOT of >>>> traction lately, and based on a few conversations in slack, looks like Kudu >>>> team is not tracking competition that much and I think there are awesome >>>> features in Snowflake that would be great to have in Kudu. >>>> >>>> I think they do support upserts now and back in-time queries - while >>>> they do not upsert/delete records, they just create new micro-partitions >>>> and it is a metadata operation - that's who they can see data in a given >>>> time. The idea of virtual warehouse is also very appealing especially in >>>> healthcare as we often need to share data with other departments or >>>> partners. >>>> >>>> do not get me wrong, I decided to use Kudu for the same reasons and I >>>> really did not have a choice on CDH other than HBase but I get questioned >>>> lately why we should not just give up CDH and go all in on Cloud with Cloud >>>> native tech like Redshift and Snowflake. It is getting harder to explain >>>> why :) >>>> >>>> My biggest gripe with Kudu besides known limitations is the management >>>> of partitions and compaction process. For our largest tables, we just max >>>> out the number of tablets as Kudu allows per cluster. Smaller tables though >>>> is a challenge and more like an art while I feel there should be an >>>> automated process with good defaults like in other data storage systems. No >>>> one expects to assign partitions manually in Snowflake, BigQuery or >>>> Redshift. No one is expected to tweak compaction parameters and deal with >>>> fast degraded performance over time on tables with a high number of >>>> deletes/updates. We are happy with performance but it is not all about >>>> performance. >>>> >>>> We also struggle on Impala side as it feels it is limiting what we can >>>> do with Kudu and it feels like Impala team is always behind. pyspark is >>>> another problem - Kudu pyspark client is lagging behind java client and >>>> very difficult to install/deploy. Some api is not even available in >>>> pyspark. >>>> >>>> And honestly, I do not like that Impala is the only viable option with >>>> Kudu. We are forced a lot of time to do heavy ETL in Hive which is like a >>>> tank but we cannot do that anymore with Kudu tables and struggle with disk >>>> spilling issues, Impala choosing bad explain plans and forcing us to use >>>> straight_join hint many times etc. >>>> >>>> Sorry if this post came a bit on a negative side, I really like Kudu >>>> and the dev team behind it rocks. But I do not like their industry is going >>>> and why Kudu is not getting the traction it deserves. >>>> >>>> On Tue, Mar 10, 2020 at 9:03 PM Cliff Resnick <cre...@gmail.com> wrote: >>>> >>>>> This is a good conversation but I don't think the comparison with >>>>> Snowflake is a fair one, at least from an older version of Snowflake (In >>>>> my >>>>> last job, about 5 years ago, I pretty much single-handedly scale tested >>>>> Snowflake in exchange for a sweetheart pricing deal) . Though Snowflake is >>>>> closed source, it seems pretty clear the architectures are quite >>>>> different. >>>>> Snowflake has no primary key index, no UPSERT capability, features that >>>>> make Kudu valuable for some use cases. >>>>> >>>>> It also seems to me that their intended workloads are quite different. >>>>> Snowflake is great for intensive analytics on demand, and can handle >>>>> deeply >>>>> nested data very well, where Kudu can't handle that at all. Snowflake is >>>>> not designed for heavy concurrency, but complex query plans for a small >>>>> group of users. If you select an x-large Snowflake cluster it's probably >>>>> because you have a large amount of data to churn through, not because you >>>>> have a large number of users. Or, at least that's how we used it. >>>>> >>>>> At my current workplace we use Kudu/Impala to handle about 30-60 >>>>> concurrent queries. I agree that getting very fussy about partitioning can >>>>> be a pain, but for the large fact tables we generally use a simple >>>>> strategy >>>>> of twelves: 12 hash x a12 (month) ranges in two 12-node clusters fronted >>>>> by a load balancer. We're in AWS, and thanks to Kudu's replication we can >>>>> use "for free" instance-store NVMe. We also have associated >>>>> compute-oriented stateless Impala Spot Fleet clusters for HLL and other >>>>> compute oriented queries. >>>>> >>>>> The net result blows away what we had with RedShift at less than 1/3 >>>>> the cost, with performance improvements mostly from better concurrency >>>>> handling. This is despite the fact that RedShift has built-in cache. We >>>>> also use streaming ingestion which, aside from being impossible with >>>>> RedShift, removes the added cost of staging. >>>>> >>>>> Getting back to Snowflake, there's no way we could use it the same way >>>>> we use Kudu, and even if we could, the cost would would probably put us >>>>> out >>>>> of business! >>>>> >>>>> On Tue, Mar 10, 2020, 10:59 AM Boris Tyukin <bo...@boristyukin.com> >>>>> wrote: >>>>> >>>>>> thanks Andrew for taking your time responding to me. It seems that >>>>>> there are no exact recommendations. >>>>>> >>>>>> I did look at scaling recommendations but that math is extremely >>>>>> complicated and I do not think anyone will know all the answers to plug >>>>>> into that calculation. We have no control really what users are doing, >>>>>> how >>>>>> many queries they run, how many are hot vs cold etc. It is not realistic >>>>>> IMHO to expect that knowledge of user query patterns. >>>>>> >>>>>> I do like the Snowflake approach than the engine takes care of >>>>>> defaults and can estimate the number of micro-partitions and even >>>>>> repartition tables as they grow. I feel Kudu has the same capabilities as >>>>>> the design is very similar. I really do not like to pick random number of >>>>>> buckets. Also we manager 100s of tables, I cannot look at them each one >>>>>> by >>>>>> one to make these decisions. Does it make sense? >>>>>> >>>>>> >>>>>> On Mon, Mar 9, 2020 at 4:42 PM Andrew Wong <aw...@cloudera.com> >>>>>> wrote: >>>>>> >>>>>>> Hey Boris, >>>>>>> >>>>>>> Sorry you didn't have much luck on Slack. I know partitioning in >>>>>>> general can be tricky; thanks for the question. Left some thoughts >>>>>>> below: >>>>>>> >>>>>>> Maybe I was not asking a clear question. If my cluster is large >>>>>>>> enough in my example above, should I go with 3, 9 or 18 tablets? or >>>>>>>> should >>>>>>>> I pick tablets to be closer to 1Gb? >>>>>>>> And a follow-up question, if I have tons of smaller tables under 5 >>>>>>>> million rows, should I just use 1 partition or still break them on >>>>>>>> smaller >>>>>>>> tablets for concurrency? >>>>>>> >>>>>>> >>>>>>> Per your numbers, this confirms that the partitions are the units of >>>>>>> concurrency here, and that therefore having more and having smaller >>>>>>> partitions yields a concurrency bump. That said, extending a scheme of >>>>>>> smaller partitions across all tables may not scale when thinking about >>>>>>> the >>>>>>> total number of partitions cluster-wide. >>>>>>> >>>>>>> There are some trade offs with the replica count per tablet server >>>>>>> here -- generally, each tablet replica has a resource cost on tablet >>>>>>> servers: WALs and tablet-related metadata use a shared disk (if you can >>>>>>> put >>>>>>> this on an SSD, I would recommend doing so), each tablet introduces some >>>>>>> Raft-related RPC traffic, each tablet replica introduces some >>>>>>> maintenance >>>>>>> operations in the pool of background operations to be run, etc. >>>>>>> >>>>>>> Your point about scan concurrency is certainly a valid one -- there >>>>>>> have been patches for other integrations that have tackled this to >>>>>>> decouple >>>>>>> partitioning from scan concurrency (KUDU-2437 >>>>>>> <https://issues.apache.org/jira/browse/KUDU-2437> and KUDU-2670 >>>>>>> <https://issues.apache.org/jira/browse/KUDU-2670> are an example, >>>>>>> where Kudu's Spark integration will split range scans into >>>>>>> smaller-scoped >>>>>>> scan tokens to be run concurrently, though this optimization hasn't made >>>>>>> its way into Impala yet). I filed KUDU-3071 >>>>>>> <https://issues.apache.org/jira/browse/KUDU-3071> to track what I >>>>>>> think is left on the Kudu-side to get this up and running, so that it >>>>>>> can >>>>>>> be worked into Impala. >>>>>>> >>>>>>> For now, I would try to take into account the total sum of resources >>>>>>> you have available to Kudu (including number of tablet servers, amount >>>>>>> of >>>>>>> storage per node, number of disks per tablet server, type of disk for >>>>>>> the >>>>>>> WAL/metadata disks), to settle on roughly how many tablet replicas your >>>>>>> system can handle (the scaling guide >>>>>>> <https://kudu.apache.org/docs/scaling_guide.html> may be helpful >>>>>>> here), and hopefully that, along with your own SLAs per table, can help >>>>>>> guide how you partition your tables. >>>>>>> >>>>>>> confused why they say "at least" not "at most" - does it mean I >>>>>>>> should design it so a tablet takes 2Gb or 3Gb in this example? >>>>>>> >>>>>>> >>>>>>> Aiming for 1GB seems a bit low; Kudu should be able to handle in the >>>>>>> low tens of GB per tablet replica, though exact perf obviously depends >>>>>>> on >>>>>>> your workload. As you show and as pointed out in documentation, larger >>>>>>> and >>>>>>> fewer tablets can limit the amount of concurrency for writes and reads, >>>>>>> though we've seen multiple GBs works relatively well for many use cases >>>>>>> while weighing the above mentioned tradeoffs with replica count. >>>>>>> >>>>>>> It is recommended that new tables which are expected to have heavy >>>>>>>> read and write workloads have at least as many tablets as tablet >>>>>>>> servers. >>>>>>>> >>>>>>> >>>>>>> if I have 20 tablet servers and I have two tables - one with 1MM >>>>>>>> rows and another one with 100MM rows, do I pick 20 / 3 partitions for >>>>>>>> both >>>>>>>> (divide by 3 because of replication)? >>>>>>> >>>>>>> >>>>>>> The recommendation here is to have at least 20 logical partitions >>>>>>> per table. That way, a scan were to touch a table's entire keyspace, the >>>>>>> table scan would be broken up into 20 tablet scans, and each of those >>>>>>> might >>>>>>> land on a different tablet server running on isolated hardware. For a >>>>>>> significantly larger table into which you expect highly concurrent >>>>>>> workloads, the recommendation serves as a lower bound -- I'd recommend >>>>>>> having more partitions, and if your data is naturally time-oriented, >>>>>>> consider range-partitioning on timestamp. >>>>>>> >>>>>>> On Sat, Mar 7, 2020 at 7:13 AM Boris Tyukin <bo...@boristyukin.com> >>>>>>> wrote: >>>>>>> >>>>>>>> just saw this in the docs but it is still confusing statement >>>>>>>> No Default Partitioning >>>>>>>> Kudu does not provide a default partitioning strategy when creating >>>>>>>> tables. It is recommended that new tables which are expected to have >>>>>>>> heavy >>>>>>>> read and write workloads have at least as many tablets as tablet >>>>>>>> servers. >>>>>>>> >>>>>>>> >>>>>>>> if I have 20 tablet servers and I have two tables - one with 1MM >>>>>>>> rows and another one with 100MM rows, do I pick 20 / 3 partitions for >>>>>>>> both >>>>>>>> (divide by 3 because of replication)? >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Sat, Mar 7, 2020 at 9:52 AM Boris Tyukin <bo...@boristyukin.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> hey guys, >>>>>>>>> >>>>>>>>> I asked the same question on Slack on and got no responses. I just >>>>>>>>> went through the docs and design doc and FAQ and still did not find an >>>>>>>>> answer. >>>>>>>>> >>>>>>>>> Can someone comment? >>>>>>>>> >>>>>>>>> Maybe I was not asking a clear question. If my cluster is large >>>>>>>>> enough in my example above, should I go with 3, 9 or 18 tablets? or >>>>>>>>> should >>>>>>>>> I pick tablets to be closer to 1Gb? >>>>>>>>> >>>>>>>>> And a follow-up question, if I have tons of smaller tables under 5 >>>>>>>>> million rows, should I just use 1 partition or still break them on >>>>>>>>> smaller >>>>>>>>> tablets for concurrency? >>>>>>>>> >>>>>>>>> We cannot pick the partitioning strategy for each table as we need >>>>>>>>> to stream 100s of tables and we use PK from RBDMS and need to come >>>>>>>>> with an >>>>>>>>> automated way to pick number of partitions/tablets. So far I was >>>>>>>>> using 1Gb >>>>>>>>> rule but rethinking this now for another project. >>>>>>>>> >>>>>>>>> On Tue, Sep 24, 2019 at 4:29 PM Boris Tyukin < >>>>>>>>> bo...@boristyukin.com> wrote: >>>>>>>>> >>>>>>>>>> forgot to post results of my quick test: >>>>>>>>>> >>>>>>>>>> Kudu 1.5 >>>>>>>>>> >>>>>>>>>> Table takes 18Gb of disk space after 3x replication >>>>>>>>>> >>>>>>>>>> Tablets Tablet Size Query run time, sec >>>>>>>>>> 3 2Gb 65 >>>>>>>>>> 9 700Mb 27 >>>>>>>>>> 18 350Mb 17 >>>>>>>>>> [image: image.png] >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> On Tue, Sep 24, 2019 at 3:58 PM Boris Tyukin < >>>>>>>>>> bo...@boristyukin.com> wrote: >>>>>>>>>> >>>>>>>>>>> Hi guys, >>>>>>>>>>> >>>>>>>>>>> just want to clarify recommendations from the doc. It says: >>>>>>>>>>> >>>>>>>>>>> https://kudu.apache.org/docs/kudu_impala_integration.html#partitioning_rules_of_thumb >>>>>>>>>>> >>>>>>>>>>> Partitioning Rules of Thumb >>>>>>>>>>> <https://kudu.apache.org/docs/kudu_impala_integration.html#partitioning_rules_of_thumb> >>>>>>>>>>> <https://kudu.apache.org/docs/kudu_impala_integration.html#partitioning_rules_of_thumb> >>>>>>>>>>> >>>>>>>>>>> - >>>>>>>>>>> >>>>>>>>>>> For large tables, such as fact tables, aim for as many >>>>>>>>>>> tablets as you have cores in the cluster. >>>>>>>>>>> - >>>>>>>>>>> >>>>>>>>>>> For small tables, such as dimension tables, ensure that each >>>>>>>>>>> tablet is at least 1 GB in size. >>>>>>>>>>> >>>>>>>>>>> In general, be mindful the number of tablets limits the >>>>>>>>>>> parallelism of reads, in the current implementation. Increasing the >>>>>>>>>>> number >>>>>>>>>>> of tablets significantly beyond the number of cores is likely to >>>>>>>>>>> have >>>>>>>>>>> diminishing returns. >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> I've read this a few times but I am not sure I understand it >>>>>>>>>>> correctly. Let me use concrete example. >>>>>>>>>>> >>>>>>>>>>> If a table ends up taking 18Gb after replication (so with 3x >>>>>>>>>>> replication it is ~9Gb per tablet if I do not partition), should I >>>>>>>>>>> aim for >>>>>>>>>>> 1Gb tablets (6 tablets before replication) or should I aim for 500Mb >>>>>>>>>>> tablets if my cluster capacity allows so (12 tablets before >>>>>>>>>>> replication)? >>>>>>>>>>> confused why they say "at least" not "at most" - does it mean I >>>>>>>>>>> should >>>>>>>>>>> design it so a tablet takes 2Gb or 3Gb in this example? >>>>>>>>>>> >>>>>>>>>>> Assume that I have tons of CPU cores on a cluster... >>>>>>>>>>> Based on my quick test, it seems that queries are faster if I >>>>>>>>>>> have more tablets/partitions...In this example, 18 tablets gave me >>>>>>>>>>> the best >>>>>>>>>>> timing but tablet size was around 300-400Mb. But the doc says "at >>>>>>>>>>> least >>>>>>>>>>> 1Gb". >>>>>>>>>>> >>>>>>>>>>> Really confused what the doc is saying, please help >>>>>>>>>>> >>>>>>>>>>> Boris >>>>>>>>>>> >>>>>>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Andrew Wong >>>>>>> >>>>>>
