SideInput is sort of like broadcast in Spark, meaning you are caching data to a local worker machines for fast lookup to reduce network/shuffle overhead. It is logical to think limit to how much memory you can have should fit in heap. In Dataflow documentation, it says limit is 20K shard. What does this mean? How big is a shard?
To answer your original question, you can configure the amount of in-memory caching done by a Dataflow worker via the --workerCacheSizeMb option on the command line, which is setWorkerCacheSizeMb if you are invoking a pipeline programmatically. The default is 100Mb.
Related
At a high / theoretical level I know exactly the type of architecture I want to build and how it would work, but I'm attempting to construct this as cheaply as possible using AWS services and my lack of familiarity with the offerings of AWS has me running in circles.
The Data
We run a video streaming platform. On busy nights we have about 100 simultaneous live streams going with upwards of 30,000 viewers. We expect this number to rise to 100,000 in the next few years. A live stream lasts, on average, 2 hours.
We send a heartbeat from our player every 10 seconds with information about the viewer -- how much data they've viewed, how much data they've buffered, what quality they're streaming, etc.
These heartbeats are sent directly to an AWS Kinesis endpoint.
Finally, we want to retain all past messages for at least 5 years (hopefully longer) so that we can look at historic analytics.
Some back of the envelope calculations suggest we will have 0.1 * 60 * 60 * 2 * 100000 * 365 * 5 = 131 billion heartbeat messages five years from now.
Our Old Pipeline
Our old system had a single Kinesis consumer. Aggregate data was stored in DynamoDB. Whenever a message arrived we would read the record from DynamoDB, update the record, then write the new record back. This read-update-write loop limited the speed at which we could process messages and made it so that each message coming in was dependent on the messages before it, so they could not be processed in parallel.
Part of the reason for this setup is that our message schema was not well designed from the outset. We send the timestamp at which the message was sent, but we do not send "amount of video watched since last heartbeat". As a result in order to compute the total viewer time we need to look up the last heartbeat message sent by this player, subtract the timestamps, and add that value. Similar issues exist with many other metrics.
Our New Pipeline
We've begun to run into scaling issues. During our peak hours analytics can be delayed by as much as four hours while waiting for a backlog of messages to be processed. If this backlog reaches 24 hours Kinesis will start deleting data. So we need to fix our pipeline to remove this dependency on past messages so we can process them in parallel.
The first part of this was updating the messages sent by our players. Our new specification includes only metrics that can be trivially sum'd with no subtraction. So we can just keep adding to the "time viewed" metric, for instance, without any regard to past messages.
The second part of this was ensuring that Kinesis never backs up. We dump the raw messages to S3 as quickly as they arrive with no processing (Kinesis Data Fire Hose) so that we can crunch analytics on them at our leisure.
Finally, we now want to actually extract information from these analytics as quickly as possible. This is where I've hit a snag.
The Questions We Want to Answer
As this is an analytics pipeline, our questions mostly revolve around filtering these messages and then aggregating fields for the remaining messages (possibly, in fact likely, with grouping). For instance:
How many Android users watched last night's stream in HD? (FILTER by stream and OS)
What's the average bandwidth usage among all users? (SUM and COUNT, with later division of the final aggregates which could be done on the dashboard side)
What percent of users last year were on any Apple device (iOS, tvOS, etc)? (COUNT, grouped by OS)
What's the average time spent buffering among Android users for streams in the past year? (a mix of all of the above)
Options
AWS Athena would allow us to query the data in S3 directly as if it were an ANSI SQL table. However reading up on Athena, unless the data is properly formatted it can be incredibly slow. Some benchmarks I've seen show that processing 1.1 billion rows of CSV data can take up to 2 minutes. I'm looking at processing 100x that much data
AWS EMR and AWS Redshift sound like they are built for this purpose, but are complicated to set up and have a high base cost to run (requiring an EC2 cluster to remain active at all times). AWS Redshift also requires data be loaded into it, which sounds like it might be a very slow process, delaying our access to analytics
AWS Glue sounds like it may be able to take the raw messages as they arrive in S3 and convert them to Parquet files for more rapid querying via Athena
We could run a job to regularly batch messages to reduce the total number that must be processed. While a stream is live we'll receive one message every 10 seconds, but we really only care about the totals for a given viewer. This means that when a 2-hour stream concludes we can combine the 720 messages we've received from that player into a single "summary" message about the viewer's experience during the whole stream. This would massively reduce the amount of data we need to process, but exactly how and when to trigger this process isn't clear to me
The Ideal Architecture
This is a Big Data problem. The generic solution to Big Data problems is "don't take your data to your query, take your query to your data". If these messages were spread across 100 small storage nodes then each node could filter, sum, and count the subset of data they hold and pass these aggregates back to a central node which sums the sums and sums the counts. If each node is only operating on 1/100th of the data set then this kind of processing could theoretically be incredibly fast.
My Confusion
While I have a theoretical understanding of the "ideal" architecture, it's not clear to me if AWS works this way or how to construct a system that will function well like this.
S3 is a black box. It's not clear if Athena queries are run on individual nodes and aggregates are further reduced elsewhere, or if there's a system reading all of the data and aggregating it in a central location
Redshift requires the data by copied into a Redshift database. This doesn't sound fast, nor distributed
It's unclear to me how EMR works or if it will suit my purpose. Still researching
AWS Glue seems like it may need to be triggered by some event?
Parquet files seems to be like CSVs, where multiple records reside in a single file. Meanwhile I'm dumping one record per file. But perhaps there's a way to fix that? e.g. batching files every minute or every 5 minutes?
RDS or a similar service might be really good for this (indexing and whatnot) but would require a guaranteed schema (or necessitate migrating if our message schema changed) which is a concern. Migrating terabytes of data if we change our message schema sounds out of the question
Finally, along with wanting to get analytics results in as "real time" as possible (ideally we want to know within 1 minute when someone joins or leaves a stream), we want the dashboards to load quickly. Waiting 30 seconds to see the count of live viewers is horrendous. Dashboards should load in 2 seconds or less (ideally)
The plan is to use QuickSight to create dashboards (our old system had a hack-y Django app that read from our DynamoDB aggregates table, but I'd like to avoid creating more code for people to maintain)
I expect you are going to get a lot of different answers and opinions from the broad set of experts you have pinged with this. There is likely no single best answer to this as there are a lot of variables. Let me give you my best advice based on my experience in the field.
Kinesis to S3 is a good start and not moving data more than needed is the right philosophy.
You didn't mention Kinesis Data Analytics and this could be a solution for SOME of your needs. It is best for questions about what is happening in the data feed right now. The longer timeframe questions are better suited for the tools you mention. If you aren't too interested in what is happening in the past 10 minutes (or so) it could be good to omit.
S3 organization will be key to performing any analytic directly on the data there. You mention parquet formatting which is good but partitioning is far more powerful. Organizing the S3 data into "days" or "hours" of data and setting up the partitioning based on this can greatly speed up any query that is limited in the amount of time that is needed (don't read what you don't need).
Important safety note on S3 - S3 is an object store and as such there is overhead for each object you reference. Having many small objects (10,000+) treated as a single set of data is going to be slow no matter what solution you go with. You need to fix this before you go forward with any solution. You see it takes upwards of .5 sec to look up an object in S3 but if the file is small the transfer time is next to nothing. Now multiply .5 sec times all the objects you have and see how long it will take to read them. This is not a function of the downstream tool you choose but of the S3 organization you have. S3 objects as part of a Big Data solution should be at least 100M in size to not suffer greatly from the object lookup time. The choice of parquet or CSV files is mute without addressing object size and partitioning first.
Athena is good for occasional queries especially if the date ranges are limited. Is this the query pattern you expect? As you say "move the compute to the data" but if you use Athena to do large cross-sectional analytics where a large percentage of the data needs to be used, you are just moving the data to Athena every time you execute this query. Don't stop thinking about data movement at the point it is stored - think about the data movements to do the analytics also.
So a big question is how much data is needed and how often to support your analytics workloads and BI functions? This is the end result you are looking for. If a high percentage of the data is needed frequently then a warehouse solution like Redshift with the data loaded to disk is the right answer. The data load time to Redshift is quite fast as it parallel loads the data from S3 (you see S3 is a cluster and Redshift is a cluster and parallel loads can be done). If loading all your data into Redshift is what you need then the load time is not your main concern - the cost is. Big powerful tool with a price tag to match. The new RA3 instance type bends this curve down significantly for large data size clusters so could be a possibility.
Another tool you haven't mentioned is Redshift Spectrum. This brings several powerful technologies together that could be important to you. First is the power of Redshift with the ability to choose smaller cluster sizes that normally would be used for your data size. S3 filtering and aggregation technology allows Spectrum to perform actions on the data in S3 (yes initial compute actions of the query are performed inside of S3 potentially greatly reducing the data moved to Redshift). If your query patterns support this data reduction in S3 then the data movement will be small and the Redshift cluster can be small (cheap) too. This can be a powerful compromise point for IoT solutions like yours since complex data models and joining are not needed.
You bring up Glue and conversion to parquet. These can be good to do but as I mentioned before partitioning of the data in S3 is usually far more powerful. The value of parquet will increase as the width of your data increases. Parquet is a columnar format so it is advantaged if only a subset of "columns" are needed. The downside is the conversion time/cost and the loss of easy human readability (which can be huge during debug).
EMR is another choice you mention but I generally advise clients against going with EMR unless they need the flexibility it brings to the analytics and they have the skills to use it well. Without these EMR tends to be an unneeded costs sink.
If this is really going to be a Big Data solution then RDS (and Aurora) not good choices. They are designed for transactional workloads, not analytics. The data size and analytics will not fit well or be cost effective.
Another tool in the space is S3 Select. Not likely what you are looking for but something to remember exists and can be a tool in the toolbox.
Hybrid solutions are common in this space if there are variable needs based on some factor. A common one "is time of day" - no one is running extensive reports at 3am so the needed performance is much less. Another is user group - some groups need simple analytics while others need much more power. Another factor is timeliness of data - does everyone need "up to the second" information or is daily information sufficient? Trying to have one tool that does everything for everybody, all the time is often a path to an expensive, oversized solution.
Since Redshift Spectrum and Athena can point at the same S3 data (well organized since both will benefit) both tools can coexist on the same data. Also, Redshift is ideal for sifting through huge mounds of data, it is ideal for producing summary tables and then writing them (in partitioned parquet) to S3 for tools like Athena to use. All these cloud services can be run on schedules and this includes Redshift and EMR (Athena is query on demand) so they don't need to run all the time. Redshift with Spectrum can run a few hours a day to perform deep analytics and summarize data for writing to S3. Your data scientist can also use Redshift for their hardcore work while Athena supports dashboards using the daily summary data and Kinesis Data Analytics as source.
Lastly you bring up a 2 sec requirement for dashboards. This is definitely possible with Quicksight backed up by Redshift or Athena but won't be met for arbitrarily complex / data intensive queries. To meet this you will need the engine to have enough horsepower to produce the data in question. Redshift with local data storage is likely the fastest (Redshift Spectrum with some data pruning done in S3 wins in some cases) and Athena is the weakest / slowest. But the power doesn't matter if the work is small - see your query workload will be a huge deciding factor. The fastest will be to load the needed data into Quicksight storage (SPICE) but this is another localized / summarized version of the data so timeliness is again a factor (how often is this updated).
Based on designing similar systems and a bunch of guesses as to what you need I'd recommend that you:
Fix your object size (Kineses can be configured to do this)
Partition your data by day
Set up a small Redshift cluster (4 X dc2.large) and use Spectrum source address the data
Connect Quicksight to Redshift
Measure the performance (and cost) and compare to requirements (there will likely be gaps)
Adjust to solution (summary tables to S3, Athena, SPICE etc.) to meet goals
The alternative is to hire someone who has set up such systems before and have them review the requirements in detail and make a less "guess-based" recommendation.
I would look into Druid. Not an AWS offering, but easily runs on AWS, with good integration with S3 and Kinesis.
Capable of reading from Kinesis, at high speeds, and make the data available for querying right away. Can also flatten and transform the data as it reads it.
Capable of doing rollups/aggregation/compaction during ingestion (and further reduce data in an async manner). From what you wrote, it seems to me that it could easily reduce the number of rows in the DB by a very large factor.
Capable of fast queries, using standard SQL.
Smart partitioning of the data to scan only the relevant dates.
The down-side is that you will need to keep a cluster up and running for ingestion and for querying. It is pretty scalable, so you can start small.
On the up-side - you're not using 10 different technologies (Athena/Glue/EMR/etc.)
You might want to consider contacting Imply, which can ease the deployment.
A usual approach a lot of companies take is they do heavy weight lifting in athena or bigquery (or some other distributed sql environment) -> aggregate intermediate results into multiple indexed+partitioned postgres/mysql/redshift/clickhouse tables and then connect their APIs to read on those tables. Of course, this works fine except the fact that with an increased amount of intermediate-aggregated data, table indices grow and problems like cumulative sum or sorting become less and less efficient.
With your problem in hand, I think you can get a lot of help with AWS Lambda. AWS Lambda provides a very feasible serverless approach towards solving large granular data problems (if used correctly). For instance, assume that your pipelines partitions incoming stream by YYYYMMMDDHHMM and stores it into some S3 path which has a Lambda listening to it (as a trigger function) then your data ingest + aggregation becomes pretty much simultaneous processes. As soon as a minute is up, a new instance of the same Lambda function will be taking care of data landing into partition YYYYMMMDDHHMM+1. So, this way, you can run thousands of simultaneous processes with a good bunch of Lambda functions doing the same thing in parallel. Of course, this is a rough picture, but I think it can greatly help.
In Dask distributed documentation, they have the following information:
For example Dask developers use this ability to build in data locality
when we communicate to data-local storage systems like the Hadoop File
System. When users use high-level functions like
dask.dataframe.read_csv('hdfs:///path/to/files.*.csv') Dask talks to
the HDFS name node, finds the locations of all of the blocks of data,
and sends that information to the scheduler so that it can make
smarter decisions and improve load times for users.
However, it seems that the get_block_locations() was removed from the HDFS fs backend, so my question is: what is the current state of Dask regarding to HDFS ? Is it sending computation to nodes where data is local ? Is it optimizing the scheduler to take into account data locality on HDFS ?
Quite right, with the appearance of arrow's HDFS interface, which is now preferred over hdfs3, the consideration of block locations is no longer part of workloads accessing HDFS, since arrow's implementation doesn't include the get_block_locations() method.
However, we already wanted to remove the somewhat convoluted code which made this work, because we found that the inter-node bandwidth on test HDFS deployments was perfectly adequate that it made little practical difference in most workloads. The extra constrains on the size of the blocks versus the size of the partitions you would like in-memory created an additional layer of complexity.
By removing the specialised code, we could avoid the very special case that was being made for HDFS as opposed to external cloud storage (s3, gcs, azure) where it didn't matter which worker accessed which part of the data.
In short, yes the docs should be updated.
I am using the dataproc cluster for spark processing. I am new to whole google cloud stuff. In our application we have 100s of jobs which uses dataproc. With every job we spawn new cluster and terminate it once the job is over. I am using pyspark for processing purpose.
Is it safe to use hybrid of stable node and pre-emptible nodes for the cost reduction?
What is the best software configuration for improving the performance of the dataproc cluser. I am aware of the in-house infrastructure optimisation of hadoop/spark cluster. Is it applicable as it is for dataroc cluster or something else is needed?
Which instance type is best suit for dataproc cluster when we are processing avro formatted data around 150GB of size.
I have tried spark's dataframe caching / persist for time optimization. But it was not that useful. Is there any way to instruct spark that entire resources (memory, processing power) belong to this job so that it can process it faster?
Does reading and writing back to GCS bucket have a performance hit? If yes, is there any way to optimize it?
Any help in time and price optimisation is appreciated. Thanks in advance.
Thanks
Manish
Is it safe to use hybrid of stable node and pre-emptible nodes for the cost reduction?
That's absolutely fine. We've used that on 300+ node clusters, only issues were with long-running clusters when nodes were getting preempted, and jobs were not optimised to account for node reclamation (no RDD replication, huge long-running DAGs). Also Tez does not like preemptible nodes getting reclaimed.
Is it applicable as it is for dataroc cluster or something else is needed?
Correct. However Google Storage driver has different characteristics when it comes to operation latency (for example, FileOutputCommitter can take huge amounts of time when trying to do recursive move or remove with overpartitioned output), and memory usage (writer buffers are 64 Mb vs 4 Kb on HDFS).
Which instance type is best suit for dataproc cluster when we are processing avro formatted data around 150GB of size.
Only performance tests can help with that.
I have tried spark's dataframe caching / persist for time optimization. But it was not that useful. Is there any way to instruct spark that entire resources (memory, processing power) belong to this job so that it can process it faster?
Make sure to use dynamic allocation and your cluster is sized to your workload. Scheduling tab in YARN UI should show utilisation close to 100% (if not, your cluster is oversized to the job, or you have not enough partitions). In Spark UI, better to have number running tasks close to number of cores (if not, it again might be not enough partitions, or cluster is oversized).
Does reading and writing back to GCS bucket have a performance hit? If yes, is there any way to optimize it?
From throughput perspective, GCS is not bad, but it is much worse in case of many small files, both from reading (when computing splits) and writing (when FileOutputCommitter) perspective. Also many parallel writes can result in OOMs due to bigger write buffer size.
My apache beam scio dataflow job is asking for more workers than my current quota. The job completes successfully, but is limited to 575 workers. What are the consequences of not giving it the RAM it is asking for. More disk IO of intermediate steps? Slower sink IO? Does it depend on what's going on with the job? In particular, my job is pretty simple really has 2 steps:
-aggregateByKey
-DO IO per key
I can run my own experiments, but I'm also interested in the cost of the job, since it isn't extremely time sensitive operation (aka I'm okay letting it run longer if it is cheaper)...
In this case, your job will have a higher runtime than if your quota was higher, but the aggregate amount of time spent performing work by all workers should be about the same.
Dataflow bills you on the amount of time each CPU, memory and storage unit is allocated. If the total CPU-hours, RAM GB-hours and storage GB-hours are about the same, your job should cost about the same.
Note: Dataflow also charges by the amount of bytes shuffled if you use the shuffle service. This should also not be affected by the number of workers.
Can someone explain using the word count example, why Spark would be faster than Map Reduce?
bafna's answer provides the memory-side of the story, but I want to add other two important facts：DAG and ecosystem
Spark uses "lazy evaluation" to form a directed acyclic graph (DAG) of consecutive computation stages. In this way, the execution plan can be optimized, e.g. to minimize shuffling data around. In contrast, this should be done manually in MapReduce by tuning each MR step. (It would be easier to understand this point if you are familiar with the execution plan optimization in RDBMS or the DAG-style execution of Apache Tez)
Spark ecosystem has established a versatile stack of components to handle SQL, ML, Streaming, Graph Mining tasks. But in the hadoop ecosystem you have to install other packages to do these individual things.
And I want to add that, even if your data is too big for main memory, you can still use spark by choosing to persist you data on disk. Although by doing this you give up the advantages of in-memory processing, you can still benefit from the DAG execution optimization.
Some informative answers on Quora:
here and here.
I think there are three primary reasons.
The main two reasons stem from the fact that, usually, one does not run a single MapReduce job, but rather a set of jobs in sequence.
One of the main limitations of MapReduce is that it persists the full dataset to HDFS after running each job. This is very expensive, because it incurs both three times (for replication) the size of the dataset in disk I/O and a similar amount of network I/O. Spark takes a more holistic view of a pipeline of operations. When the output of an operation needs to be fed into another operation, Spark passes the data directly without writing to persistent storage. This is an innovation over MapReduce that came from Microsoft's Dryad paper, and is not original to Spark.
The main innovation of Spark was to introduce an in-memory caching abstraction. This makes Spark ideal for workloads where multiple operations access the same input data. Users can instruct Spark to cache input data sets in memory, so they don't need to be read from disk for each operation.
What about Spark jobs that would boil down to a single MapReduce job? In many cases also these run faster on Spark than on MapReduce. The primary advantage Spark has here is that it can launch tasks much faster. MapReduce starts a new JVM for each task, which can take seconds with loading JARs, JITing, parsing configuration XML, etc. Spark keeps an executor JVM running on each node, so launching a task is simply a matter of making an RPC to it and passing a Runnable to a thread pool, which takes in the single digits of milliseconds.
Lastly, a common misconception probably worth mentioning is that Spark somehow runs entirely in memory while MapReduce does not. This is simply not the case. Spark's shuffle implementation works very similarly to MapReduce's: each record is serialized and written out to disk on the map side and then fetched and deserialized on the reduce side.