1. Use shallow trees and the clever workaround presented in the article.
2. Don't use Spark for tasks that require complex logic.
People should trace out the line of reasoning that leads them to use tools like Spark. It is convoluted and contingent - it goes back to work done at Google in the early 2000s, when the key to getting good price / performance was using a large number of commodity machines. Because they were cheap, these machines would break often, so you needed some really smart fault tolerance technology like Hadoop/HDFS, which was followed by Spark.
The current era is completely different. Now the key to good price / performance is to light up machines on-demand and then shut them down, only paying for what you use - and perhaps using the spot market. You don't need to worry about storage - that's taken care of by the cloud provider, and you can't "bring the computation to the data" like in the old days, removing one of the big advantages of Hadoop/HDFS. Because they are doing mostly IO and networking, and because computers are just more resilient nowadays, jobs rarely fail because of hardware errors. So almost the entire rationale that led to Hadoop/HDFS/Spark is gone. But people still use Spark - and put up with "accidentally exponential behavior" - because the tech industry is do dominated by groupthink and marketing dollars.
Edit: I don’t mean to suggest that there is no reason to use Spark, but ~95% of the usage in industry is unnecessary now and should be avoided.
The most common reason for spark use today is ETL+DataLakes (ie., cloud object stores and ETL in/out).
It seems actual analysis is happening in fast databases that receive data from the object stores.
can anyone here comment on this paradigm?
1. https://www.usenix.org/system/files/conference/hotos15/hotos...
No, Flink is much better.
But we ended up implementing a single level Multi-AND [2] so that this no longer a tree for just AND expressions & can be vectorized neater than the nested structure with a function call for each (this looks more like a tail-call rather than a recursive function).
The ORC CNF conversion has a similar massively exponential item inside which is protected by a check for 256 items or less[3].
[1] - https://github.com/t3rmin4t0r/captain-hook/blob/master/src/m...
[2] - https://issues.apache.org/jira/browse/HIVE-11398
[3] - https://github.com/apache/hive/blob/master/storage-api/src/j...
p1 AND (p2 AND (p3 AND notp4)) -> (p1 AND p2) AND (p3 AND notp4)
ALL (p1 p2 p3 NOT(p4))
That needs an "in SQL", the standard imperative language operator ordering has short-cut operations in there (a is null or a.value == true) etc.
In the code I work with, this actually sorts the conditions based on estimated selectivity[1] and type (long compares to constant are cheaper on a columnar data-set due to the SIMD, but string isn't etc).
> Is there any reason to choose binary trees for this?
The parse-tree does come off as binary because inserting logical parentheses makes it easier to tackle, because there are association rules which neatly go into a BinaryOp structure when dealing with operator precedence in parsing.
So it is easier to handle the parsing when you treat (a + (b + c) ) and (a / (b / c)) in similar fashion.
I won't make the same mistake again if I have to build a SQL engine, but this actually made the logical expression match the parse tree very closely and was a good enough generalization until the traversal time bugs[2] started to pop up.
[1] - https://github.com/apache/hive/blob/master/common/src/java/o... [2] - https://issues.apache.org/jira/browse/HIVE-9166
val transformedLeftTemp = transform(tree.left)
val transformedLeft = if (transformedLeftTemp.isDefined) {
transformedLeftTemp
} else NoneThe real implementation has a mutable `builder` argument used to gradually build the converted filter. If we perform the `transform().isDefined` call directly on the "main" builder, but the subtree turns out to not be convertible, we can mess up the state of the builder.
The second example from the post would look roughly like this:
val transformedLeft = if (transform(tree.left, new Builder()).isDefined) {
transform(tree.left, mainBuilder)
} else None
There's a more detailed explanation in the old comment to the method: https://github.com/apache/spark/pull/24068/files#diff-5de773... .
val transformedLeft = transform(tree.left)The rhetorical question saying that surely that weird refactor of two different functions into one, followed by calling that new, non-trivial function twice for no reason surely shouldn't affect performance.. He already lost me during the premise of the article.
Then, ne writes some code that works around the library bug by modifying the input losslessly into something that's more easily processed by the library.
Finally, ne patches the library bug and shares the patch.
All of this is also kinda fucking obvious to not just me, but a lot of people, so I'm having a really hard time grasping if you've mixed up the illustrative simplification with the actual code, or if you think that the best engineering approach is to always patch your environment bugs instead of modifying your input, or if you just don't have a Github account or for some other reason can't read the patch.
Between that patch and https://github.com/apache/spark/pull/24910 you can see why the code is what it is.
Supposing spark is your ETL machinery... would it not make more sense to ETL this into a database?
Our main Spark workload is pretty spiky. We have low load during most of the day, and very high load at certain times - either system-wide, or because a large customer triggered an expensive operation. Using Spark as our distributed query engine allows us to quickly spin up new worker nodes and process the high load in a timely manner. We can then downsize the cluster again to keep our compute spend in check.
And just to provide some context on our data size, here's an article about how we use Citus at Heap - https://www.citusdata.com/customers/heap . We store close to a petabyte of data in our distributed Citus cluster. However, we've found Spark to be significantly better at queries with large result sets - our Connect product syncs a lot of data from our internal storage to customers' warehouses.
When you want to process N+1 TB/PB its hard to throw standard relational approaches at it imo.
SQL is strings all the way down, testing the database itself is often shitshow...