Context:
My text input pipeline currently consists of two main parts:
I. A complex text preprocessing and exporting of tf.SequenceExamples to tfrecords (custom tokenization, vocabulary creation, statistics calculation, normalization and many more over the full dataset as well as per each individual example). That is done once for each data configuration.
II. A tf.Dataset (TFRecords) pipeline that does quite a bit of processing during training, too (string_split into characters, table lookups, bucketing, conditional filtering, etc.).
Original Dataset is present across multiple locations (BigQuery, GCS, RDS, ...).
Problem:
The problem is that as the production dataset increases rapidly (several terabytes), it is not feasible to recreate a tfrecords files for each possible data configuration (part 1 has a lot of hyperparameters) as each will have an enormous size of hundreds of terabytes. Not to mention, that tf.Dataset reading speed surprisingly slows down when tf.SequenceExamples or tfrecords grow in size.
There are quite a few possible solutions:
Apache Beam + Cloud DataFlow + feed_dict;
tf.Transform;
Apache Beam + Cloud DataFlow + tf.Dataset.from_generator;
tensorflow/ecosystem + Hadoop or Spark
tf.contrib.cloud.BigQueryReader
, but neither of the following seem to fully fulfill my requierments:
Streaming and processing on the fly data from BigQuery, GCS, RDS, ... as in part I.
Sending data (protos?) directly to tf.Dataset in one way or another to be used in part II.
Fast and reliable for both training and inference.
(optional) Being able to pre-calculate some full pass statistics over the selected part of the data.
EDIT: Python 3 support would be just wonderful.
What is the most suitable choice for the tf.data.Dataset pipeline? What are the best practices in this case?
Thanks in advance!
I recommend to orchestrate the whole use case using Cloud Composer(GCP integration of Airflow).
Airflow provided operators which let you orchestrate a pipeline with a script.
In your case you can use the dataflow_operator to have the dataflow job spin up when you have enough data to process.
To get the data from BigQuery you can use the bigquery_operator.
Furthermore you can use the python operator or the bash operator to monitor and pre-calculate statistics.
Related
I have terabytes of data stored in S3 bucket in parquet format. I want to develop prototype followed by scaling. The data is really huge. Everyday, it is 50gb incremental. It is structured conventional data (is not image, video or audio). The history for prototype isn't decided yet but if it is 3 months, then it will be 90days x 50gb = 4500gb for prototype or (9000 gb for 6 months).
I want to do data processing, derive some new variables, EDA followed by modelling (feature engineering & implement unsupervised deep learning algorithms). Can anyone suggest me the best way here? For e.g. Use sagemaker notebook, write data processing python scripts there, save processed data to S3 folder, and then apply algorithms? or use EMR for data processing followed by SageMaker for EDA+Modeling. Or any other best way?
There are couple of ways depending on the finer details. You could use Data Wrangler for EDA or SageMaker Processing with ShardedByS3Key for scaling out. Post EDA you can use SageMaker built-in algorithms or develop your own and SageMaker Pipelines for CI/CD
We have large volumes (10 to 400 billion) of raw data in BigQuery tables. We have a requirement to process this data to convert and create the data in the form of star schema tables (probably a different dataset in bigquery) which can then be accessed by atscale.
Need pros and cons between two options below:
1. Write complex SQL within BigQuery which reads data form source dataset and then loads to target dataset (used by Atscale).
2. Use PySpark or MapReduce with BigQuery connectors from Dataproc and then load the data to BigQuery target dataset.
The complexity of our transformations involve joining multiple tables at different granularity, using analytics functions to get the required information, etc.
Presently this logic is implemented in vertica using multiple temp tables for faster processing and we want to re-write this processing logic in GCP (Big Query or Data Proc)
I went successfully with option 1: Big Query is very capable to run the very complex transformation with SQL, on top of that you can also run them incrementally with time range decorators. Note that it takes a lot of time and resources to take data back and forth to BigQuery. When running BigQuery SQL data never leaves BigQuery in the first place and you already have all raw logs there. So as long your problem can be solved by a series of SQL I believe this is the best way to go.
We moved out Vertica reporting cluster, rewriting successfully ETL last year, with option 1.
Around a year ago, I've written POC comparing DataFlow and series of BigQuery SQL jobs orchestrated by potens.io workflow allowing SQL parallelization at scale.
I took a good month to write DataFlow in Java with 200+ data points and complex transformation with terrible debugging capability at a time.
And a week to do the same using a series of SQL with potens.io utilizing
Cloud Function for Windowed Tables and parallelization with clustering transient tables.
I know there's been bunch improvement in CloudDataFlow since then, but at a time
the DataFlow did fine only at a million scale and never-completed at billions record input (main reason shuffle cardinality went little under billions of records, with each records having 200+ columns). And the SQL approach produced all required aggregation under 2 hours for a dozen billion. Debugging and easiest of troubleshooting with potens.io helped a lot too.
Both BigQuery and DataProc can handle huge amounts of complex data.
I think that you should consider two points:
Which transformation would you like to do in your data?
Both tools can make complex transformations but you have to consider that PySpark will provide you a full programming language processing capability while BigQuery will provide you SQL transformations and some scripting structures. If only SQL and simple scripting structures can handle your problem, BigQuery is an option. If you need some complex scripts to transform your data or if you think you'll need to build some extra features involving transformations in the future, PySpark may be a better option. You can find the BigQuery scripting reference here
Pricing
BigQuery and DataProc have different pricing systems. While in BigQuery you'd need to concern about how much data you would process in your queries, in DataProc you have to concern about your cluster's size and VM's configuration, how much time your cluster would be running and some other configurations. You can find the pricing reference for BigQuery here and for DataProc here. Also, you can simulate the pricing in the Google Cloud Platform Pricing Calculator
I suggest that you create a simple POC for your project in both tools to see which one has the best cost benefit for you.
I hope these information help you.
I already have terabytes of data stored on BigQuery and I'm wondering to perform heavy data transformations on it.
Considering COSTS and PERFORMANCE, what the best approach you guys would suggest to perform these transformations for future usage of these data on BigQuery?
I'm considering a few options:
1. Read raw data from DataFlow and then load the transformed data back into BigQuery?
2. Do it directly from BigQuery?
Any ideas about how to proceed with this?
I wrote down some most important things about performance, you can find there consideration regarding your question about using DataFlow.
Best practices considering performance:
Choosing file format:
BigQuery supports a wide variety of file formats for data ingestion. Some are going to be naturally faster than others. When optimizing for load speed, prefer using the AVRO file format, which is binary, row-based format and enables to split it and then read it in parallel with multiple workers.
Loading data from compressed files, specifically CSV and JSON, is going to be slower than loading data in a other format. And the reason being is because, since the compression of Gzip is not splitable, there is a need to take that file, load it onto a slot within BQ, and then do the decompression, and then finally parallelize the load afterwards.
**FASTER**
Avro(Compressed)
Avro(Uncompressed)
Parquet/ORC
CSV
JSON
CSV (Compressed)
JSON(Compressed
**SLOWER**
ELT / ETL:
After loading data into BQ, you can think about transformations (ELT or ETL). So in general, you want to prefer ELT over ETL where possible. BQ is very scalable and can handle large transformations on a ton of data. ELT is also quite a bit simpler, because you could just write some SQL queries, transform some data and then move data around between tables, and not have to worry about managing a separate ETL application.
Raw and staging tables:
Once, you have started loading data into BQ, in general, within your warehouse, you're going to want to leverage raw and staging tables before publishing to reporting tables. The raw table essentially contains the full daily extract, or a full load of the data that they're loading. Staging table then is basically your change data capture table, so you can utilize queries or DML to marge that data into your staging table and have a full history of all the data that's being inserted. And then finally your reporting tables are going to be the ingest that you publish out to your users.
Speeding up pipelines using DataFlow:
When you're getting into streaming loads really complex batch loads (that doesn't really fit into SQL cleanly), you can leverage DataFlow or DataFusion to speed up those pipelines, and do more complex activities on that data. And if you're starting with streaming, I recommend using the DataFlow templates - Google provides it for loading data from multiple different places and moving data around. You can find those templates in DataFlow UI, within Create Job from Template button, you'll find all these templates.
And if you find that it mostly fits your use case, but want to make one slight modification, all those templates are also open sourced (so you can go to repo, modify the code to fit your needs).
Partitioning:
Partition in BQ physically split your data on disk, based on ingestion time or based on a column within your data. Efficiently query over the parts of the table you want. This provides huge cost and performance benefits, especially on large fact tables. Whenever you have a fact table or temporal table, utilize a partition column on your date dimension.
Cluster Frequently Accessed Fields:
Clustering allows you to physically order data within a partition. So you can do Clustering by one or multiple keys. This provide massive performance benefits when used properly.
BQ reservations:
It allows to create reservations of slots, assign project to those reservations, so you can allocate more or less resources to certain types of queries.
Best practices considering saving costs you can find in official documentation.
I hope it helps you.
According to this Google Cloud Documentation, the following questions should be done to choose between DataFlow or BigQuery tool for ELT.
Although the data is small and can quickly be uploaded by using the BigQuery UI, for the purpose of this tutorial you can also use Dataflow for ETL. Use Dataflow for ETL into BigQuery instead of the BigQuery UI when you are performing massive joins, that is, from around 500-5000 columns of more than 10 TB of data, with the following goals:
You want to clean or transform your data as it's loaded into BigQuery, instead of storing it and joining afterwards. As a result,
this approach also has lower storage requirements because data is only
stored in BigQuery in its joined and transformed state.
You plan to do custom data cleansing (which cannot be simply achieved with SQL).
You plan to combine the data with data outside of the OLTP, such as logs or remotely accessed data, during the loading process.
You plan to automate testing and deployment of data-loading logic using continuous integration or continuous deployment (CI/CD).
You anticipate gradual iteration, enhancement, and improvement of the ETL process over time.
You plan to add data incrementally, as opposed to performing a one-time ETL.
I have a multiple stage ETL transform stage using pandas. Basically, I load almost 2Gb of data from Mongodb and then I apply several functions in the columns. My question is if there's any way to break those transformations in multiple Airflow tasks.
The options I have considered are:
Creating a temporary table in Mongodb and loading/storing the transformed data frame between steps. I found this cumbersome and totally prone to a non-usual overhead due to disk I/O
Passing data among the tasks using XCom. I think this is a nice solution but I worry about the sheer size of the data. The docs explicitly state
Any object that can be pickled can be used as an XCom value, so users should make sure to use objects of appropriate size.
Using an in-memory storage between steps. Maybe saving the data in a Redis server or something, but I'm not really sure if that would be any better than just using XCom altogether.
So, does any of you have any tips on how to handle this situation? Thanks!
I've been developing Spark jobs for some years using on-premise clusters and our team recently moved to the Google Cloud Platform allowing us to leverage the power of BigQuery and such.
The thing is, I now often find myself writing processing steps in SQL more than in PySpark since it is :
easier to reason about (less verbose)
easier to maintain (SQL vs scala/python code)
you can run it easily on the GUI if needed
fast without having to really reason about partitioning, caching and so on...
In the end, I only use Spark when I've got something to do that I can't express using SQL.
To be clear, my workflow is often like :
preprocessing (previously in Spark, now in SQL)
feature engineering (previously in Spark, now mainly in SQL)
machine learning model and predictions (Spark ML)
Am I missing something ?
Is there any con in using BigQuery this way instead of Spark ?
Thanks
A con I can see is the additional time required by the Hadoop cluster to create and finish the job. By making a direct request to BigQuery, this extra time can be decreased.
If your tasks need parallel processing, I would recommend using Spark, but if your app is mainly used to access to BQ, you might want to use the BQ Client Libraries and separate your current tasks:
BigQuery Client Libraries. They are optimized to connect to BQ. Here is a QuickStart and you can use different programming languages like python or java, among others.
Spark jobs. If you still need to perform transformations in Spark and need to read the data from BQ you can use the Dataproc-BQ connector. While this connector is installed in Dataproc by default, you can install it on-premises so that you can continue running you SparkML jobs with BQ data. Just in case it helps, you might want to consider using some GCP services like AutoML, BQ ML, AI Platform Notebooks, etc., they are specialized services for Machine Learning and AI.
I'm using PySpark (on GCP Dataproc), BigQuery and we have jobs in both. I will summarize my vision about Pros and Cons of one system against the other. And I do admit that your environment could be different, so that something which I think is Pros might not be like this for you.
Pros of Spark:
better testing of the code, simpler to build unit tests and run them with mocked data and classes, rather in trying to do this with BigQuery
it's possible to use SQL (SparkSQL) for operations and even combine operations over different data sources (DB, files, BQ)
we have JSON files in the format which is not valid for BigQuery, and it cannot parse them (while files have valid JSON format)
possible to implement naturally more complicated logic for some cases, for example, traversing arrays in nested fields and other complicated calculations
better custom monitoring is possible, when we need to check specific metrics in the pipeline we can send related metrics (StatsD, etc.) easier
more natural for CI/CD processes
Pros of BigQuery (all with a note: if all data is available):
simplicity of SQL, when all data is available in a convenient format
DBAs who are not familiar with Python/Scala still could contribute (bcs they know SQL)
awesome infrastructure behind the scene, very performant
With both approaches it's possible to check quickly the result in GUI. For example, Jupyter Notebook allows to run PySpark instantly. I cannot add my notes about ML related traits, though.