There's a 64kb limit on strings in Kudu but otherwise you can store any binary data in them.   https://docs.cloudera.com/documentation/kudu/5-10-x/topics/kudu_known_issues.html#schema_design_limitations ... View more

Upgrading to a newer version of Impala will solve most scalability issues that you'd see on Impala 2.9, mostly because of https://blog.cloudera.com/scalability-improvement-of-apache-impala-2-12-0-in-cdh-5-15-0/. ... View more

The versions of Apache Impala in Cloudera are "based on" Apache Impala releases but will often include substantial additional features and fixes. We do a lot of work beyond just repackaging the upstream releases. Please consult the Cloudera release notes for the version of CDH/CDP you're using if you want to understand what features are present. Just looking at the version string won't give you that information.   CDH 6.3.4 includes that fix. See https://docs.cloudera.com/documentation/enterprise/6/release-notes/topics/rg_cdh_634_fixed_issues.html   Most new features have been going into CDP only - we have been rolling out new Impala features in CDP public cloud on a fairly continual basis and then these are trickling down into CDP private cloud releases.   Some limited features have gone into minor CDH release - i.e. 6.3.0, but we've generally been prioritizing stability there and focusing on CDP. We shipped a preview version of the remote read cache in CDH 6.3.0 - https://docs.cloudera.com/documentation/enterprise/6/release-notes/topics/rg_cdh_630_new_features.html#impala_new_630 . Ranger support is not in CDH - the general idea there is to allow migrating from Sentry to Ranger as part of the CDH->CDP upgrade process. ... View more

We have some background on schema evolution in Parquet in the docs - https://docs.cloudera.com/runtime/7.2.2/impala-reference/topics/impala-parquet.html. See "Schema Evolution for Parquet Tables". Some of the details are specific to Impala but the concepts are the same across engines including Hive and Spark that use parquet tables. At a high level, you can think of the data files being immutable while the table schema evolves. If you add a new column at the end of the table, for example, that updates the table schema but leaves the parquet files unchanged. When the table is queried, the table schema and parquet file schema are reconciled and the new column's values will be all NULL.   If you want to modify the existing rows and include new non-NULL values, that would require rewriting the data, e.g. with an INSERT OVERWRITE statement for a partition or a CREATE TABLE .. AS SELECT to create an entirely new table. Keep in mind that traditional Parquet tables are not optimized for workloads with updates - Apache Kudu in particular and also transactional tables in Hive3+ have support for row-level updates that is more convenient/efficient.   We definitely don't require rewriting the whole table every time you want to add a column, that would be impractical for large tables! ... View more

"... or because memory is running low or because the extra scanner threads are not needed to keep up with the consumer from the scan node." is how I should've finished that. Not sure what happened there, I did hit the main points but truncated the last sentence. That part of Impala has had a lot of improvements for performance and observability (i.e. more info in the profile) since CDH5.10, FWIW, I'd guess on a later version this wouldn't be a problem or would be easier to debug at least. ... View more

One difference is how fast it's reading from disk - i.e. TotalRawHdfsReadTime. In CDH5.12 that includes both time spend fetching metadata from the HDFS namenode and actually reading the data off disk. If you're saying that it's only slow on one node, that probably rules out HDFS namenode slowness, which is a common cause. So probably it's actually slower doing the I/O. Note: in CDH5.15 we split out the namenode RPC time into TotalRawHdfsOpenTime to make it easier to debug things like this. I don't know exactly why I/O would be slower on that one node, it might require inspecting the host to see what's happening and if there's more CPU or I/O load on that host.   We've seen that happen if a node is more heavily loaded than other nodes because of some kind of uneven data distribution. E.g. one file is very frequently accessed, maybe if there's a dimension table that is referenced in many queries. That can sometimes be addressed by setting SCHEDULE_RANDOM_REPLICA as a query hint or query option https://docs.cloudera.com/documentation/enterprise/latest/topics/impala_hints.html or https://docs.cloudera.com/documentation/enterprise/latest/topics/impala_schedule_random_replica.html. Or even by enabling HDFS caching for the problematic table (HDFS caching spreads load across all cached replicas). Another possible cause, based on that profile, is that it's competing for scanner threads with other queries running on the same node - AverageScannerThreadConcurrency is lower in the slow case. This can either be because other concurrent queries grabbed scanner threads first (there's a global soft limit of 3x # cpus per node) or because ... View more

Are you sure there isn't a stack trace for IllegalStateException in the impala daemon logs or catalog daemon logs? That would help match it to a bug. ... View more

You need to run compute stats on the base tables referenced by the views - compute stats directly on a view isn't supported. ... View more