Databricks Community

Good summary! yes those are the main strategies I can think of.

You do not have to cache anything to make it work. You would decide that based on whether you want to spend memory/storage to avoid recomputing the DataFrame, like when you may use it in multiple operations afterwards.

With Z-order and Hilbert curves, the effectiveness of clustering decreases with each column added - so you'd want to zorder only the columns that you's actually use so that it's speed up your workloads.

coalesce avoids a full shuffle and could be used to decrease the number of partitionsrepartition results in a full shuffle and could be used to increase or decrease the number of partitions

We recommend having a top level folder to run jobs against. Best practice detailed here: https://docs.databricks.com/repos.html#best-practices-for-integrating-repos-with-cicd-workflows

We don't currently support git submodules.

If you have a private git server (e.g. behind VPN, IP whitelist), you will need to be enrolled in the git proxy private preview to use Repos, please contact your account team.

import sqlContext.implicits._import org.apache.spark.sql.functions._inputDF.select(   unix_timestamp($"unix_timestamp").alias("unix_timestamp"),   from_utc_timestamp($"unix_timestamp".cast(DataTypes.TimestampType), "UTC").alias("UTC"),   from_utc_tim...

While Spark does use a micro-batch execution model, this does not have much impact on applications, because the batches can be as short as 0.5 seconds. In most applications of streaming big data, the analytics is done over a larger window (say 10 min...