Spark SQL .distinct() performance - sql

I want to source few hundreds of gigabytes from a database via JDBC and then process it using Spark SQL. Currently I am doing some partitioning at that data and process is by batches of milion records. The thing is that I would like also to apply some deduplication to my dataframes and I was going to leave that idea of separated batches processing and try to process those hundreds of gigabytes using a one dataframe partitioned accordingly.
The main concern is: how will .distinct() work in such case? Will Spark SQL firstly try to load ALL the data into the RAM and then apply deduplication involving many shuffles and repartitioning? Do I have to ensure that a cluster has enough of RAM to contain that raw data or maybe it will be able to help itself with HDD storage (thus killing the performance)?
Or maybe I should do it without Spark - move the data to the target storage and there apply distinct counts and detect duplicates and get rid off them?

Spark SQL does NOT use predicate pushdown for distinct queries; meaning that the processing to filter out duplicate records happens at the executors, rather than at the database. So, your assumption regarding shuffles happening over at the executors to process distinct is correct.
Inspite of this, I would still advise you to go ahead and perform the de-duplication on Spark, rather than build a separate arrangement for it. My personal experience with distinct has been more than satisfactory. It has always been the joins which push my buttons.

Related

BigQueryIO Read vs fromQuery

Say in Dataflow/Apache Beam program, I am trying to read table which has data that is exponentially growing. I want to improve the performance of the read.
BigQueryIO.Read.from("projectid:dataset.tablename")
or
BigQueryIO.Read.fromQuery("SELECT A, B FROM [projectid:dataset.tablename]")
Will the performance of my read improve, if i am only selecting the required columns in the table, rather than the entire table in above?
I am aware that selecting few columns results in the reduced cost. But would like to know the read performance in above.
You're right that it will reduce cost instead of referencing all the columns in the SQL/query. Also, when you use from() instead of fromQuery(), you don't pay for any table scans in BigQuery. I'm not sure if you were aware of that or not.
Under the hood, whenever Dataflow reads from BigQuery, it actually calls its export API and instructs BigQuery to dump the table(s) to GCS as sharded files. Then Dataflow reads these files in parallel into your pipeline. It does not ready "directly" from BigQuery.
As such, yes, this might improve performance because the amount of data that needs to be exported to GCS under the hood, and read into your pipeline will be less i.e. less columns = less data.
However, I'd also consider using partitioned tables, and then even think about clustering them too. Also, use WHERE clauses to even further reduce the amount of data to be exported and read.

How to store millions of statistics records efficiently?

We have about 1.7 million products in our eshop, we want to keep record of how many views this products had for 1 year long period, we want to record the views every atleast 2 hours, the question is what structure to use for this task?
Right now we tried keeping stats for 30 days back in records that have 2 columns classified_id,stats where stats is like a stripped json with format date:views,date:views... for example a record would look like
345422,{051216:23212,051217:64233} where 051216,051217=mm/dd/yy and 23212,64233=number of views
This of course is kinda stupid if you want to go 1 year back since if you want to get the sum of views of say 1000 products you need to fetch like 30mb from the database and calculate it your self.
The other way we think of going right now is just to have a massive table with 3 columns classified_id,date,view and store its recording on its own row, this of course will result in a huge table with hundred of millions of rows , for example if we have 1.8 millions of classifieds and keep records 24/7 for one year every 2 hours we need
1800000*365*12=7.884.000.000(billions with a B) rows which while it is way inside the theoritical limit of postgres I imagine the queries on it(say for updating the views), even with the correct indices, will be taking some time.
Any suggestions? I can't even imagine how google analytics stores the stats...
This number is not as high as you think. In current work we store metrics data for websites and total amount of rows we have is much higher. And in previous job I worked with pg database which collected metrics from mobile network and it collected ~2 billions of records per day. So do not be afraid of billions in number of records.
You will definitely need to partition data - most probably by day. With this amount of data you can find indexes quite useless. Depends on planes you will see in EXPLAIN command output. For example that telco app did not use any indexes at all because they would just slow down whole engine.
Another question is how quick responses for queries you will need. And which steps in granularity (sums over hours/days/weeks etc) for queries you will allow for users. You may even need to make some aggregations for granularities like week or month or quarter.
Addition:
Those ~2billions of records per day in that telco app took ~290GB per day. And it meant inserts of ~23000 records per second using bulk inserts with COPY command. Every bulk was several thousands of records. Raw data were partitioned by minutes. To avoid disk waits db had 4 tablespaces on 4 different disks/ arrays and partitions were distributed over them. PostreSQL was able to handle it all without any problems. So you should think about proper HW configuration too.
Good idea also is to move pg_xlog directory to separate disk or array. No just different filesystem. It all must be separate HW. SSDs I can recommend only in arrays with proper error check. Lately we had problems with corrupted database on single SSD.
First, do not use the database for recording statistics. Or, at the very least, use a different database. The write overhead of the logs will degrade the responsiveness of your webapp. And your daily backups will take much longer because of big tables that do not need to be backed up so frequently.
The "do it yourself" solution of my choice would be to write asynchronously to log files and then process these files afterwards to construct the statistics in your analytics database. There is good code snippet of async write in this response. Or you can benchmark any of the many loggers available for Java.
Also note that there are products like Apache Kafka specifically designed to collect this kind of information.
Another possibility is to create a time series in column oriented database like HBase or Cassandra. In this case you'd have one row per product and as many columns as hits.
Last, if you are going to do it with the database, as #JosMac pointed, create partitions, avoid indexes as much as you can. Set fillfactor storage parameter to 100. You can also consider UNLOGGED tables. But read thoroughly PostgreSQL documentation before turning off the write-ahead log.
Just to raise another non-RDBMS option for you (so a little off topic), you could send text files (CSV, TSV, JSON, Parquet, ORC) to Amazon S3 and use AWS Athena to query it directly using SQL.
Since it will query free text files, you may be able to just send it unfiltered weblogs, and query them through JDBC.

Efficiently perform COUNT DISTINCT with spark, on csvs?

I have a large volume of data, and I'm looking to efficiently (ie, using a relatively small Spark cluster) perform COUNT and DISTINCT operations one of the columns.
If I do what seems obvious, ie load the data into a dataframe:
df = spark.read.format("CSV").load("s3://somebucket/loadsofcsvdata/*").toDF()
df.registerView("someview")
and then attempt to run a query:
domains = sqlContext.sql("""SELECT domain, COUNT(id) FROM someview GROUP BY domain""")
domains.take(1000).show()
my cluster just crashes and burns - throwing out of memory exceptions or otherwise hanging/crashing/not completing the operation.
I'm guessing that somewhere along the way there's some sort of join that blows one of the executors' memory?
What's the ideal method for performing an operation like this, when the source data is at massive scale and the target data isn't (the list of domains in the above query is relatively short, and should easily fit in memory)
related info available at this question: What should be the optimal value for spark.sql.shuffle.partitions or how do we increase partitions when using Spark SQL?
I would suggest to tune your executors settings. Especially, setting following parameters correctly can provide dramatic improvement in performance.
spark.executor.instances
spark.executor.memory
spark.yarn.executor.memoryOverhead
spark.executor.cores
In your case, I would also suggest to tune Number of partitions, especially bump up following param from default 200 to higher value, as per requirement.
spark.sql.shuffle.partitions

How to Let Spark Handle Bigger Data Sets?

I have a very complex query that needs to join 9 or more tables with some 'group by' expressions . Most of these tables have almost the same of numbers of the rows. These tables also have some columns that can be used as the 'key' to partition the tables.
Previously, the app ran fine, but now the data set has 3~4 times data as before. My tests turned out if the row count of each table is less than 4,000,000, the application can still run pretty nicely. However, if the count is more than that, the application writes hundreds of terabytes of shuffling and the application stalls (no matter how I adjust the memory, partition, executors, etc.). The actual data probably is just dozens of Gs.
I would think that if the partitioning works properly, Spark shouldn't do shuffle so much and the join should be done on each node. It is puzzling that why Spark is not so 'smart' to do so.
I could split the data set (with the 'key' I mentioned above) into many data sets that these data sets can be dealt with independently. But the burden will be on myself...it discounts the very reason to use Spark. What other approaches that could help?
I use Spark 2.0 over Hadoop YARN.
My tests turned out if the row count of each table is less than 4,000,000, the application can still run pretty nicely. However, if the count is more than that, the application writes hundreds of terabytes of shuffling
When joining datasets if the size of one side is less than a certain configurable size, spark broadcasts the entire table to each executor so that join may be performed locally everywhere. Your above observation is consistent with this. You can also provide broadcast hint explicitly to the spark, like so df1.join(broadcast(df2))
Other than that, can you please provide more specifics about your problem?
[Sometime ago I was also grappling with the issue of join and shuffles for one of our jobs that had to handle couple of TBs. We were using RDDs (and not the dataset api). I wrote about my findings [here]1. These may be of some use to you are try to reason about the underlying data shuffle.]
Update: According to documentation -- spark.sql.autoBroadcastJoinThreshold is the configurable property key. 10 MB is its default value. And it does the following:
Configures the maximum size in bytes for a table that will be broadcast to all worker nodes when performing a join. By setting this value to -1 broadcasting can be disabled. Note that currently statistics are only supported for Hive Metastore tables where the command ANALYZE TABLE COMPUTE STATISTICS noscan has been run.
So apparently, this is supported only for the Hive tables.

Improve apache hive performance

I have 5GB of data in my HDFS sink. When I run any query on Hive it takes more than 10-15 minutes to complete. The number of rows I get when I run,
select count(*) from table_name
is 3,880,900. My VM has 4.5 GB mem and it runs on MBP 2012. I would like to know if creating index in the table will have any performance improvement. Also are there any other ways to tell hive to only use this much amount of data or rows so as to get results faster? I am ok even if the queries are run for a lesser subset of data at least to get a glimpse of the results.
Yes, indexing should help. However, getting a subset of data (using limit) isn't really helpful as hive still scans the whole data before limiting the output.
You can try using RCFile/ORCFile format for faster results. In my experiments, RCFile based tables executed queries roughly 10 times faster than textfile/sequence file based tables.
Depending on the data you are querying you can get gains by using the different file formats like ORC, Parquet. What kind of data are you querying, is it structured or unstructured data? What kind of queries are you trying to perform? If it is structured data you can see gains also by using other SQL on Hadoop solutions such as InfiniDB, Presto, Impala etc...
I am an architect for InfiniDB
http://infinidb.co
SQL on Hadoop solutions like InfiniDB, Impala and others work by you loading your data through them at which they will perform calculations, optimizations etc... to make that data faster to query. This helps tremendously for interactive analytical queries, especially when compared to something like Hive.
With that said, you are working with 5GB of data (but data always grows! someday could be TBs), which is pretty small so you can still work in the worlds of the some of the tools that are not intended for high performance queries. Your best solution with Hive is to look at how your data is and see if ORC or Parquet could benefit your queries (columnar formats are good for analytic queries).
Hive is always going to be one of the slower options though for performing SQL queries on your HDFS data. Hortonworks with their Stinger initiative is making it better, you might want to check that out.
http://hortonworks.com/labs/stinger/
The use case sounds fit for ORC, Parquet if you are interested in a subset of the columns. ORC with hive 0.12 comes with PPD which will help you discarding blocks while running the queries using the meta data that it stores for each column.
We did an implementation on top of hive to support bloom filters in the meta data indexes for ORC files which gave a performance gain of 5-6X.
What is average number of Mapper/Reducer tasks launched for the queries you execute? Tuning some parameters can definitely help.