I'm ingesting very large files into Cassandra 2.0 and I'm noticing that my ingest rate into Cassandra will be x3 slower than the rate at which I'm getting new files to ingest. Given that, and trying to avoid memory problems, what are my options for keeping up with ingest?
I was initially thinking that I could have multiple clients writing, possibly each to a different "seed" node in the cluster. If I am careful about not accessing the same file twice will that cause problems with the node I/O? What is the best way to go about doing this? Based on google searches I have seen things like batch driver statements can help, but I'm reading in CSV files which need to be cleaned first...
Related
When saving to a delta table we avoid 'df.coalesce(1)' but when saving to csv or parquet we(my team) add 'df.coalesce(1)'. Is it a common practise? Why? Is it mandatory?
In most cases when I have seen df.coalesce(1) it was done to generate only one file, for example, import CSV file into Excel, or for Parquet file into the Pandas-based program. But if you're doing .coalesce(1), then the write happens via single task, and it's becoming the performance bottleneck because you need to get data from other executors, and write it.
If you're consuming data from Spark or other distributed system, having multiple files will be beneficial for performance because you can write & read them in parallel. By default, Spark writes N files into the directory where N is the number of partitions. As #pltc noticed, this may generate the big number of files that's often not desirable because you'll get performance overhead from accessing them. So we need to have a balance between the number of files and their size - for Parquet and Delta (that is based on Parquet), having the bigger files bring several performance advantages - you read less files, you can get better compression for data inside the file, etc.
For Delta specifically, having .coalesce(1) having the same problem as for other file formats - you're writing via one task. Relying on default Spark behaviour and writing multiple files is beneficial from performance point of view - each node is writing its data in parallel, but you can get too many small files (so you may use .coalesce(N) to write bigger files). For Databricks Delta, as it was correctly pointed by #Kafels, there are some optimizations that will allow to remove that .coalesce(N) and do automatic tuning achieve the best throughput (so called "Optimized Writes"), and create bigger files ("Auto compaction") - but they should be used carefully.
Overall, the topic of optimal file size for Delta is an interesting topic - if you have big files (1Gb is used by default by OPTIMIZE command), you can get better read throughput, but if you're rewriting them with MERGE/UPDATE/DELETE, then big files are bad from performance standpoint, and it's better to have smaller (16-64-128Mb) files, so you can rewrite less data.
TL;DR: it's not mandatory, it depends on the size of your dataframe.
Long answer:
If your dataframe is 10Mb, and you have 1000 partitions for example, each file would be about 10Kb. And having so many small files would reduce Spark performance dramatically, not to mention when you have too many files, you'll eventually reach OS limitation of the number of files. Any how, when your dataset is small enough, you should merge them into a couple of files by coalesce.
However, if your dataframe is 100G, technically you still can use coalesce(1) and save to a single file, but later on you will have to deal with less parallelism when reading from it.
I am looking for a way to create a streaming application that can withstand millions of events per second and output a distinct count of those events in real time. As this stream is unbounded by any time window it obviously has to be backed by some storage. However, I cannot find the best way to do this maintaining a good level of abstraction (meaning that I want a framework to handle storing and counting for me, otherwise I don't need a framework at all). The preferred storage for me are Cassandra and Redis (both ideally).
The options I've considered are Flink, Spark and Kafka Streams. I do know the differences between them, but I still can't pick the best solution. Can someone advice? Thanks in advance.
Regardless of which solution you choose, if you can withstand it not being 100% accurate (but being very very close), you can have your operator using HyperLogLog (there are Java implementations available). This allows you to not actually have to keep around data about each individual item, drastically reducing your memory usage.
Assuming Flink, the necessary state is quite small (< 1MB), so can easily use the FSStateBackend which is heap-based and checkpoints to the file system, allowing you to reduce serialization overhead.
Again assuming you go with Flink, Using the [ContinuousEventTimeTrigger][2], you can also get a view into how many unique items are currently being tracked.
I'd suggest to reconsider the choice of storage system. Using an external system is significantly slower than using local state. Flink applications locally maintain state on the JVM heap or in RocksDB (on disk) and can checkpoint it in regular intervals to persistent storage such as HDFS. This state can grow very big (10s of TBs) and still be efficiently maintained because checkpoints can be incrementally and asynchronously done. This gives much better performance than sending a query to an external system for each record.
If you still prefer Redis or Cassandra, you can use Flink's AsyncIO operator to send asynchronous requests to improve the throughput of your application.
I am working with some legacy systems in investment banking domain, which are very unfriendly in the sense that, only way to extract data from them is through a file export/import. Lots of trading takes place and large number of transactions are stored on these system.
Q is how to read large number of large files on NFS and dump it on a system on which analytics can be done by something like Spark or Samza.
Back to issue. Due nature of legacy systems, we are extracting data and dumping into files. Each file is in hundreds of gigabyte size.
I feel next step is to read these and dump to Kafka or HDFS, or maybe even Cassandra or HBase. Reason being I need to run some financial analytics on this data. I have two questions:
How to efficiently read large number of large files which are located on one or numerous machines
Apparently you've discovered already that mainframes are good at writing large numbers of large files. They're good at reading them too. But that aside...
IBM has been pushing hard on Spark on z/OS recently. It's available for free, although if you want support, you have to pay for that. See: https://www-03.ibm.com/systems/z/os/zos/apache-spark.html My understanding is that z/OS can be a peer with other machines in a Spark cluster.
The z/OS Spark implementation comes with a piece that can read data directly from all sorts of mainframe sources: sequential, VSAM, DB2, etc. It might allow you to bypass the whole dump process and read the data directly from the source.
Apparently Hadoop is written in Java, so one would expect that it should be able to run on z/OS with little problem. However, watch out for ASCII vs. EBCDIC issues.
On the topic of using Hadoop with z/OS, there's a number of references out there, including a red piece: http://www.redbooks.ibm.com/redpapers/pdfs/redp5142.pdf
You'll note that in there they make mention of using the CO:z toolkit, which I believe is available for free.
However you mention "unfriendly". I'm not sure if that means "I don't understand this environment as it doesn't look like anything I've used before" or it means "the people I'm working with don't want to help me". I'll assume something like the latter since the former is simply a learning opportunity. Unfortunately, you're probably going to have a tough time getting the unfriendly people to get anything new up and running on z/OS.
But in the end, it may be best to try to make friends with those unfriendly z/OS admins as they likely can make your life easier.
Finally, I'm not sure what analytics you're planning on doing with the data. But in some cases it may be easier/better to move the analytics process to the data instead of moving the data to the analytics.
The simplest way to do it better is zconnector, a ibm product for data ingestion between mainframe to hadoop cluster.
I managed to find an answer. The biggest bottleneck is that reading files is essentially a serial operation.. that is the most efficient way to read from a disk. So for one file I am stuck with a single thread reading it from NFS and sending it to HDFS or Kafka via their APIs.
So it appears best way is to make sure that the source from where data is coming dumps files in multiple NFS folders. That point onward I can run multiple processes to load data to HDFS or Kafka since they are highly parallelized.
How to load? One good way is to mount NFS into Hadoop infrastructure and use distcp. There are other possiblities too which open up once we make sure files are available from large number of NFS. Otherwise remember, reading file is a serial operation. Thanks.
We have an application which the clients use to track their procurement cycle. We need to build a solution which will help the users to pull any column from any table in a particular subject area and they should be able to see all the rows of the result of this join of the tables from which the columns have been pulled. It needs to be similar to a Salesforce kind of reporting solution. We are looking at HDFS and Spark in Azure HDInsight to support these kind of querying capabilities. We would like to know if this is a valid use case for Spark. The volume of the joins of all tables can easily touch 500 million rows which will be pulled into the Spark driver memory before being displayed to the user.
Please let me know if this is something that can be done using Spark.
As per my understanding, Spark is mostly used for batch processing. If your use case is directly user-facing, then I am doubtful about using Spark because there may be better solutions(or alternate architectures). Becuase joining 500 million rows in realtime sounds crazy!
The volume of the joins of all tables can easily touch 500 million rows which will be pulled into the Spark driver memory before being displayed to the user.
This is another thing that puzzled me. Pulling all the 500 million rows into RAM of a single java process doesn't sound right because of the obvious reasons.
Updated
Just using spark for processing huge data will not be effective for realtime solutions(like your use case). But, Spark will be very effective if you are going to pre-process your data, cache the results using some other system, prepare views using the results can be served to your users. More or less similar to Lambda Architecture.
Spark on Yarn cluster to periodically process the data and generate/update the different views, a distributed storage system (preferably columnar storage systems) to cache the views, a REST API to serve the views to users.
Late reply to the question, but in case someone else is reading this in future. AWS Redshift does exactly this.
Background
We have recently started a "Big Data" project where we want to track what users are doing with our product - how often they are logging in, which features they are clicking on, etc - your basic user analytics stuff. We still don't know exactly what questions we will be asking, but most of it will be "how often did X occur over the last Y months?" type of thing, so we started storing the data sooner rather than later thinking we can always migrate, re-shape etc when we need to but if we don't store it it is gone forever.
We are now looking at what sorts of questions we can ask. In a typical RDBMS, this stage would consist of slicing and dicing the data in many different dimensions, exporting to Excel, producing graphs, looking for trends etc - it seems that for Cassandra, this is rather difficult to do.
Currently we are using Apache Spark, and submitting Spark SQL jobs to slice and dice the data. This actually works really well, and we are getting the data we need, but it is rather cumbersome as there doesn't seem to be any native API for Spark that we can connect to from our workstations, so we are stuck using the spark-submit script and a Spark app that wraps some SQL from the command line and outputs to a file which we then have to read.
The question
In a table (or Column Family) with ~30 columns running on 3 nodes with RF 2, how bad would it be to add an INDEX to every non-PK column, so that we could simply query it using CQL across any column? Would there be a horrendous impact on the performance of writes? Would there be a large increase in disk space usage?
The other option I have been investigating is using Triggers, so that for each row inserted, we populated another handful of tables (essentially, custom secondary index tables) - is this a more acceptable approach? Does anyone have any experience of the performance impact of Triggers?
Impact of adding more indexes:
This really depends on your data structure, distribution and how you access it; you were right before when you compared this process to RDMS. For Cassandra, it's best to define your queries first and then build the data model.
These guys have a nice write-up on the performance impacts of secondary indexes:
https://pantheon.io/blog/cassandra-scale-problem-secondary-indexes
The main impact (from the post) is that secondary indexes are local to each node, so to satisfy a query by indexed value, each node has to query its own records to build the final result set (as opposed to a primary key query where it is known exactly which node needs to be quired). So there's not just an impact on writes, but on read performance as well.
In terms of working out the performance on your data model, I'd recommend using the cassandra-stress tool; you can combine it with a data modeler tool that Datastax have built, to quickly generate profile yamls:
http://www.datastax.com/dev/blog/data-modeler
For example, I ran the basic stress profile without and then with secondary indexes on the default table, and the "with indexes" batch of writes took a little over 40% longer to complete. There was also an increase in GC operations / duration etc.