We have a spark job (HDInsight) and its run time is increasing slowly over the last couple of months. Based on the spark UI , are there any indicators to say that there is dataskew thats why its performance is degrading ? below is stage details, please see the median and 75th percentile difference. How should i go about optimizing this job ? appreciate any guidance
Also what is the optimal value for spark.sql.shuffle.partitions given the input dataset size is around 12 GB and cluster has got 128 cores ?
Normally we use the following approach to identify possible stragglers and you have caught on to that:
When Max duration among completed tasks is significanly higher than
Median or 75th Percentile value, then this indicates the possibility
of Stragglers.
The median in your case tells enough.
You need to salt the key to distribute the data better - which can be complicated, or look at current partitioning approach.
This article https://medium.com/swlh/troubleshooting-stragglers-in-your-spark-application-47f2568663ec provides good guidance.
Related
scikit-learn==0.21.2
Hierarchal Agglomerative Clustering algorithm response time is increasing exponentially when increasing the dataset.
My Data set is textual. Each Document is 7-10 words long.
Using the following code to perform the Clustering.
hac_model = AgglomerativeClustering(affinity=consine,
linkage=complete,
compute_full_tree=True,
connectivity=None, memory=None,
n_clusters=None,
distance_threshold=0.7)
cluster_matrix = hac_model.fit_predict(matrix)
where the matrix of size are:
5000x1500 taking 17 seconds
10000*2000 taking 113 seconds
13000*2418 taking 228 seconds
I can't control 5000, 10000, 15000 as that is the size of input, or the feature set size(i.e 1500,2000,2418) since I am using BOW model(TFIDF).
I end up using all the unique words(after removing stopwords) as my feature list. this list grows as the input size increases.
So two questions.
How do I avoid increase in feature set size irrespective of increase in the size of input data set
Is there a way I can improve on the performance of the Algorithm without compromising on the quality?
Standard AGNES hierarchical clustering is O(n³+n²d) in complexity. So the number of instances is much more a problem than the number of features.
There are approaches that typically run in O(n²d), although the worst case remains the same, so they will be much faster than this. With these you'll usually run into memory limits first... Unfortunately, this isn't implemented in sklearn for all I know, so you'll have to use other clustering tools - or write the algorithm yourself.
I'd like to get the 0.95 percentile memory usage of my pods from the last x time. However this query start to take too long if I use a 'big' (7 / 10d) range.
The query that i'm using right now is:
quantile_over_time(0.95, container_memory_usage_bytes[10d])
Takes around 100s to complete
I removed extra namespace filters for brevity
What steps could I take to make this query more performant ? (except making the machine bigger)
I thought about calculating the 0.95 percentile every x time (let's say 30min) and label it p95_memory_usage and in the query use p95_memory_usage instead of container_memory_usage_bytes, so that i can reduce the amount of points the query has to go through.
However, would this not distort the values ?
As you already observed, aggregating quantiles (over time or otherwise) doesn't really work.
You could try to build a histogram of memory usage over time using recording rules, looking like a "real" Prometheus histogram (consisting of _bucket, _count and _sum metrics) although doing it may be tedious. Something like:
- record: container_memory_usage_bytes_bucket
labels:
le: 100000.0
expr: |
container_memory_usage_bytes > bool 100000.0
+
(
container_memory_usage_bytes_bucket{le="100000.0"}
or ignoring(le)
container_memory_usage_bytes * 0
)
Repeat for all bucket sizes you're interested in, add _count and _sum metrics.
Histograms can be aggregated (over time or otherwise) without problems, so you can use a second set of recording rules that computes an increase of the histogram metrics, at much lower resolution (e.g. hourly or daily increase, at hourly or daily resolution). And finally, you can use histogram_quantile over your low resolution histogram (which has a lot fewer samples than the original time series) to compute your quantile.
It's a lot of work, though, and there will be a couple of downsides: you'll only get hourly/daily updates to your quantile and the accuracy may be lower, depending on how many histogram buckets you define.
Else (and this only came to me after writing all of the above) you could define a recording rule that runs at lower resolution (e.g. once an hour) and records the current value of container_memory_usage_bytes metrics. Then you could continue to use quantile_over_time over this lower resolution metric. You'll obviously lose precision (as you're throwing away a lot of samples) and your quantile will only update once an hour, but it's much simpler. And you only need to wait for 10 days to see if the result is close enough. (o:
The quantile_over_time(0.95, container_memory_usage_bytes[10d]) query can be slow because it needs to take into account all the raw samples for all the container_memory_usage_bytes time series on the last 10 days. The number of samples to process can be quite big. It can be estimated with the following query:
sum(count_over_time(container_memory_usage_bytes[10d]))
Note that if the quantile_over_time(...) query is used for building a graph in Grafana (aka range query instead of instant query), then the number of raw samples returned from the sum(count_over_time(...)) must be multiplied by the number of points on Grafana graph, since Prometheus executes the quantile_over_time(...) individually per each point on the displayed graph. Usually Grafana requests around 1000 points for building smooth graph. So the number returned from sum(count_over_time(...)) must be multiplied by 1000 in order to estimate the number of raw samples Prometheus needs to process for building the quantile_over_time(...) graph. See more details in this article.
There are the following solutions for reducing query duration:
To add more specific label filters in order to reduce the number of selected time series and, consequently, the number of raw samples to process.
To reduce the lookbehind window in square brackets. For example, changing [10d] to [1d] reduces the number of raw samples to process by 10x.
To use recording rules for calculating coarser-grained results.
To try using other Prometheus-compatible systems, which may process heavy queries at faster speed. Try, for example, VictoriaMetrics.
I have 2 tables in Hive: user and item and I am trying to calculate cosine similarity between 2 features of each table for a cartesian product between the 2 tables, i.e. Cross Join.
There are around 20000 users and 5000 items resulting in 100 million rows of calculation. I am running the compute using Scala Spark on Hive Cluster with 12 cores.
The code goes a little something like this:
val pairs = userDf.crossJoin(itemDf).repartition(100)
val results = pairs.mapPartitions(computeScore) // computeScore is a function to compute the similarity scores I need
The Spark job will always fail due to memory issues (GC Allocation Failure) on the Hadoop cluster. If I reduce the computation to around 10 million, it will definitely work - under 15 minutes.
How do I compute the whole set without increasing the hardware specifications? I am fine if the job takes longer to run and does not fail halfway.
if you take a look in the Spark documentation you will see that spark uses different strategies for data management. These policies are enabled by the user via configurations in the spark configuration files or directly in the code or script.
Below the documentation about data management policies:
"MEMORY_AND_DISK" policy would be good for you because if the data (RDD) does not fit in the ram then the remaining partitons will be stored in the hard disk. But this strategy can be slow if you have to access the hard drive often.
There are few steps of doing that:
1. Check the expected Data volume after cross join and divide this by 200 as spark.sql.shuffle.partitions by default comes as 200. It has to be more than 1 GB raw data to each partition.
2. Calculate each row size and multiply with another table row count , you will be able to estimated the rough Volume. The process will work much better in Parquet in comparison to CSV file
3. spark.sql.shuffle.partitions needs to be set based on Total Data Volume/500 MB
4. spark.shuffle.minNumPartitionsToHighlyCompress needs to set a little less than Shuffle Partition
5. Bucketize the source parquet data based on the joining column for both of the files/tables
6. Provide a High Spark Executor Memory and Manage the Java Heap memory too considering the heap space
df.groupBy("c1").agg(sum("n1")).distinct.count()
would take 10 seconds
df.groupBy("c1").agg(sum("n1"), sum("n2")).distinct.count()
would take 20 seconds
It suprises me since row storage of DFs. Do you have same experience & how does this make sense? Also ideas how to make 2 sums run in more similar time to 1 sum? spark 2.2.0
I don't think "agg" takes two much more time in second case. I would look towards distinct.
You're executing distinct based on extra column n2, which gives broader distribution and increase complexity of distinct calulation.
It makes sense:
You increase number of computations twofold.
You increase shuffle size roughly 50%.
Both changes will impact overall performance, even if final result is small and impact on distinct is negligible.
So, I understand that in general one should use coalesce() when:
the number of partitions decreases due to a filter or some other operation that may result in reducing the original dataset (RDD, DF). coalesce() is useful for running operations more efficiently after filtering down a large dataset.
I also understand that it is less expensive than repartition as it reduces shuffling by moving data only if necessary. My problem is how to define the parameter that coalesce takes (idealPartionionNo). I am working on a project which was passed to me from another engineer and he was using the below calculation to compute the value of that parameter.
// DEFINE OPTIMAL PARTITION NUMBER
implicit val NO_OF_EXECUTOR_INSTANCES = sc.getConf.getInt("spark.executor.instances", 5)
implicit val NO_OF_EXECUTOR_CORES = sc.getConf.getInt("spark.executor.cores", 2)
val idealPartionionNo = NO_OF_EXECUTOR_INSTANCES * NO_OF_EXECUTOR_CORES * REPARTITION_FACTOR
This is then used with a partitioner object:
val partitioner = new HashPartitioner(idealPartionionNo)
but also used with:
RDD.filter(x=>x._3<30).coalesce(idealPartionionNo)
Is this the right approach? What is the main idea behind the idealPartionionNo value computation? What is the REPARTITION_FACTOR? How do I generally work to define that?
Also, since YARN is responsible for identifying the available executors on the fly is there a way of getting that number (AVAILABLE_EXECUTOR_INSTANCES) on the fly and use that for computing idealPartionionNo (i.e. replace NO_OF_EXECUTOR_INSTANCES with AVAILABLE_EXECUTOR_INSTANCES)?
Ideally, some actual examples of the form:
Here 's a dataset (size);
Here's a number of transformations and possible reuses of an RDD/DF.
Here is where you should repartition/coalesce.
Assume you have n executors with m cores and a partition factor equal to k
then:
The ideal number of partitions would be ==> ???
Also, if you can refer me to a nice blog that explains these I would really appreciate it.
In practice optimal number of partitions depends more on the data you have, transformations you use and overall configuration than the available resources.
If the number of partitions is too low you'll experience long GC pauses, different types of memory issues, and lastly suboptimal resource utilization.
If the number of partitions is too high then maintenance cost can easily exceed processing cost. Moreover, if you use non-distributed reducing operations (like reduce in contrast to treeReduce), a large number of partitions results in a higher load on the driver.
You can find a number of rules which suggest oversubscribing partitions compared to the number of cores (factor 2 or 3 seems to be common) or keeping partitions at a certain size but this doesn't take into account your own code:
If you allocate a lot you can expect long GC pauses and it is probably better to go with smaller partitions.
If a certain piece of code is expensive then your shuffle cost can be amortized by a higher concurrency.
If you have a filter you can adjust the number of partitions based on a discriminative power of the predicate (you make different decisions if you expect to retain 5% of the data and 99% of the data).
In my opinion:
With one-off jobs keep higher number partitions to stay on the safe side (slower is better than failing).
With reusable jobs start with conservative configuration then execute - monitor - adjust configuration - repeat.
Don't try to use fixed number of partitions based on the number of executors or cores. First understand your data and code, then adjust configuration to reflect your understanding.
Usually, it is relatively easy to determine the amount of raw data per partition for which your cluster exhibits stable behavior (in my experience it is somewhere in the range of few hundred megabytes, depending on the format, data structure you use to load data, and configuration). This is the "magic number" you're looking for.
Some things you have to remember in general:
Number of partitions doesn't necessarily reflect
data distribution. Any operation that requires shuffle (*byKey, join, RDD.partitionBy, Dataset.repartition) can result in non-uniform data distribution. Always monitor your jobs for symptoms of a significant data skew.
Number of partitions in general is not constant. Any operation with multiple dependencies (union, coGroup, join) can affect the number of partitions.
Your question is a valid one, but Spark partitioning optimization depends entirely on the computation you're running. You need to have a good reason to repartition/coalesce; if you're just counting an RDD (even if it has a huge number of sparsely populated partitions), then any repartition/coalesce step is just going to slow you down.
Repartition vs coalesce
The difference between repartition(n) (which is the same as coalesce(n, shuffle = true) and coalesce(n, shuffle = false) has to do with execution model. The shuffle model takes each partition in the original RDD, randomly sends its data around to all executors, and results in an RDD with the new (smaller or greater) number of partitions. The no-shuffle model creates a new RDD which loads multiple partitions as one task.
Let's consider this computation:
sc.textFile("massive_file.txt")
.filter(sparseFilterFunction) // leaves only 0.1% of the lines
.coalesce(numPartitions, shuffle = shuffle)
If shuffle is true, then the text file / filter computations happen in a number of tasks given by the defaults in textFile, and the tiny filtered results are shuffled. If shuffle is false, then the number of total tasks is at most numPartitions.
If numPartitions is 1, then the difference is quite stark. The shuffle model will process and filter the data in parallel, then send the 0.1% of filtered results to one executor for downstream DAG operations. The no-shuffle model will process and filter the data all on one core from the beginning.
Steps to take
Consider your downstream operations. If you're just using this dataset once, then you probably don't need to repartition at all. If you are saving the filtered RDD for later use (to disk, for example), then consider the tradeoffs above. It takes experience to become familiar with these models and when one performs better, so try both out and see how they perform!
As others have answered, there is no formula which calculates what you ask for. That said, You can make an educated guess on the first part and then fine tune it over time.
The first step is to make sure you have enough partitions. If you have NO_OF_EXECUTOR_INSTANCES executors and NO_OF_EXECUTOR_CORES cores per executor then you can process NO_OF_EXECUTOR_INSTANCES*NO_OF_EXECUTOR_CORES partitions at the same time (each would go to a specific core of a specific instance).
That said this assumes everything is divided equally between the cores and everything takes exactly the same time to process. This is rarely the case. There is a good chance that some of them would be finished before others either because of locallity (e.g. the data needs to come from a different node) or simply because they are not balanced (e.g. if you have data partitioned by root domain then partitions including google would probably be quite big). This is where the REPARTITION_FACTOR comes into play. The idea is that we "overbook" each core and therefore if one finishes very quickly and one finishes slowly we have the option of dividing the tasks between them. A factor of 2-3 is generally a good idea.
Now lets take a look at the size of a single partition. Lets say your entire data is X MB in size and you have N partitions. Each partition would be on average X/N MBs. If N is large relative to X then you might have very small average partition size (e.g. a few KB). In this case it is usually a good idea to lower N because the overhead of managing each partition becomes too high. On the other hand if the size is very large (e.g. a few GB) then you need to hold a lot of data at the same time which would cause issues such as garbage collection, high memory usage etc.
The optimal size is a good question but generally people seem to prefer partitions of 100-1000MB but in truth tens of MB probably would also be good.
Another thing you should note is when you do the calculation how your partitions change. For example, lets say you start with 1000 partitions of 100MB each but then filter the data so each partition becomes 1K then you should probably coalesce. Similar issues can happen when you do a groupby or join. In such cases both the size of the partition and the number of partitions change and might reach an undesirable size.