- How is data storage handled in Pachyderm?
- What object storage backends are currently supported?
- What is version control for data?
- What are the benefits of version control for data?
- How do you guarantee I won’t lose data in Pachyderm (i.e. replication and persistence)?
- How do I get data from other sources into Pachyderm?
- How do I get data out of Pachyderm into another system?
- Does Pachyderm handle data locality?
- Where/how can I deploy Pachyderm?
- Can I use other schedulers such as Docker Swarm or Mesos?
- Can I run Pachyderm locally?
- What are containerized analytics?
- What is the data access model?
- What are jobs and how do they work?
- What are pipelines and how do they work?
- How does Pachyderm manage pipeline dependencies?
- How do I perform batched analytics in Pachyderm?
- How do I perform streaming analytics in Pachyderm?
- How is my computation parallelized?
- How does Pachyderm let me do incremental processing?
- Is there a SQL interface for Pachyderm?
- How does Pachyderm compare to Hadoop?
- How does Pachyderm compare to Spark?
- What are the major use cases for Pachyderm?
- Is Pachyderm enterprise production ready?
- How does Pachyderm handle logging?
- Does Pachyderm only work with Docker containers?
- How do I get enterprise support for Pachyderm?
- What if I find bugs or have questions about using Pachyderm?
- How do I start contributing to Pachyderm?
How is data storage handled in Pachyderm?¶
Pachyderm stores your data in any generic object storage (S3, GSC, Ceph, etc). You can link your object storage backend to Pachyderm by following our cloud deployment guide and passing your credentials as a Kubernetes secret.
What object storage backends are supported?¶
S3 and GCS are fully supported and are the recommended backends for Pachyderm. Support for Ceph and others are coming soon! Want to help us support more storage backends? Check out the GH issue!
What is version control for data?¶
We’ve all used version control for code before — Pachyderm gives you the same
semantics for petabytes of data. We even borrow our terminology from Git. In
Pachyderm, data is organized into
repos. If you want to add or change data in
a repo, you simply
commit make your changes, and then
commit. This will create an immutable snapshot of the data that you can
reference later. Just like in Git, only the diff of the data is saved so
there is no duplication. Pachyderm exposes data as a set of diffs so you can
easily view how your data has changed over time, run a job over a previous view
of your data, or revert to a known good state if something goes wrong. Finally,
Pachyderm also let’s you branch entire data sets so you can manipulate files
and explore the data without affecting anyone else’s work. Just like with
branching in Git, Pachyderm doesn’t create multiple copies of the data when you
create a branch, we just store the changes you make to it.
What are the benefits of version control for data?¶
Instant revert: If something goes wrong with your data, you can immediately revert your live cluster back to a known good state.
View diffs: Analyze how your data is changing over time.
Incrementality: Only process the new data instead of recomputing everything.
Immutable data: Run analysis written today over your data from last month.
Team collaboration: Everyone can manipulate and work on the same data without stepping on each others toes.
How do you guarantee I won’t lose data in Pachyderm (i.e. replication and persistence)?¶
Your data doesn’t actually live in Pachyderm, it stays in object storage (S3 or GCS), so it has all the safety guarantees of those underlying systems.
How do I get data from other sources into Pachyderm?¶
Pachyderm has three main methods for getting data into the system.
How do I get data out of Pachyderm into another system?¶
In addition to using the same ways you get data into the system, you can also use pipelines. Users often want to move the final results of a pipeline into another tool such as Redshift or MySQL so that it can be easily queried through BI tools. To accomplish this, it’s common to add a final stage to your pipeline which reads data from Pachyderm and writes it directly to whatever other tool you want. Redshift for example, can load data directly from an S3 bucket so the last pipeline stage can just write to that specific bucket.
Does Pachyderm handle data locality?¶
Most object stores like S3 and GCS don’t provide any notion of locality and so Pachyderm similarly can’t provide data locality in our API. In practice, we’ve generally found that data locality is not a bottleneck when optimizing for performance in modern data centers.
Where/how can I deploy Pachyderm?¶
Once you have Kubernetes running, Pachyderm is just a one line deploy. Since Pachyderm’s only dependency is Kubernetes, it can be run locally, AWS, Google Cloud, Azure, or on-prem. Check out our local installation and cloud deployment guides.
Can I use other schedulers such as Docker Swarm or Mesos?¶
Right now, Pachyderm requires Kubernetes, but we’ve purposely built it to be modular and work with the other schedulers as well. Swarm and Mesos support will be added in the future!
What are containerized analytics?¶
Rather than thinking in terms of map or reduce jobs, Pachyderm thinks in terms of pipelines expressed within a container. To process data, you simply create a containerized program which reads and writes to the local filesystem. Since everything is packaged up in a container, pipelines are easily portable, completely isolated, and simple to monitor.
What is the data access model?¶
To process data, you simply create a containerized program which reads and
writes to the local filesystem at
/pfs/.... Pachyderm will take your
container and inject data into it by way of a FUSE volume. We’ll then
automatically replicate your container, showing each copy a different chunk of
data and processing it all in parallel.
What are jobs and how do they work?¶
A job in Pachyderm is a one-off transformation or processing of data. To run a
job use the
create-job command. In Pachyderm, jobs are meant for
experimentation or exploring your data. Once you have a job that’s working well and producing useful results, you can “productionize” it by turning it into a
What are pipelines and how do they work?¶
Pipelines are data transformations that are “subscribed” to data changes on their input repos and automatically create jobs to process the new data as it comes in. A pipeline is defined by a JSON spec that describes one or more transformations to execute when new input data is committed. All the details of a pipeline spec are outlined in our documentation and demonstrated in our examples.
How does Pachyderm manage pipeline dependencies?¶
Dependencies for pipelines are handled explicitly in the pipeline spec. Pipelines output their results to a repo of the same name. The “input” for a pipeline can be any set of repos, either those containing raw data or one that was automatically created by another pipeline. For example, a pipeline stage called “filter” would create a repo also called “filter” where it would store the output data. A second pipeline called “sum” could have “filter” as an input. By this method Pachyderm, actually creates a DAG of data, not jobs. The full picture would look like this: raw data enters Pachyderm which triggers the “filter” pipeline. The “filter” pipeline outputs its results in a commit to the “filter” repo which triggers the “sum” pipeline. The final results would be available in the “sum” repo. Check out our Fruit Stand demo to see exactly this example.
How do I perform batched analytics in Pachyderm?¶
Batched analytics are the bread and butter of Pachyderm. Often times the first stage in a batched job is a database dump or some other large swath of new data entering the system. In Pachyderm, this would create a new commit on a repo which would trigger all your ETL and analytics pipelines for that data. One-off batched jobs can also be manually run on any data.
How do I perform streaming analytics in Pachyderm?¶
Streaming and batched jobs are done exactly the same way in Pachyderm. Creating a commit is an incredibly cheap operation so you can even make one commit per second if you want! By just changing the frequency of commits, you can seamlessly transition from a large nightly batch job down to a streaming operation processing tiny micro-batches of data.
How is my computation parallelized?¶
Both jobs and pipelines have a “paralellism” parameter as outlined in the pipeline spec. This parameter dictates how
many containers Pachyderm spins up to process your data in parallel. For
“paralellism”: 10 would create up to 10 containers that each process 1/10 of
the data. Each pipeline can have a different parallelization factor, giving you
fine-grain control over the utilization of your cluster.
paralellism can be set to
0 in which case Pachyderm will set it automatically based on the size of the
How does Pachyderm let me do incremental processing?¶
Pachyderm exposes all your data in diffs, meaning we show you the new data that has been added since the last time a pipeline was run. Pachyderm will smartly only process the new data and append those results to the output from the previous run. We have extensive documentation on incrementality that’ll show you the fine-grain control you can have to optimizing computation.
Is there a SQL interface for Pachyderm?¶
Not yet, but it’s coming soon! If you want to query your data using SQL, you can easily create a pipeline that pushes data from Pachyderm into your favorite SQL tool such as PostGres.
How does Pachyderm compare to Hadoop?¶
Pachyderm is inspired by the Hadoop ecosystem but shares no code with it. Instead, we leverage the container ecosystem to provide the broad functionality of Hadoop with the ease of use of Docker. Similar to Hadoop, Pachyderm offers virtually infinite horizontal scaling for both storage and processing power. That said, there are two bold new ideas in Pachyderm:
- Containers as the core processing primitive — You can do analysis using any languages or libraries you want.
- Version Control for data — We let your team collaborate effectively on data using a commit-based distributed filesystem (PFS), similar to what Git does with code.
How does Pachyderm compare to Spark?¶
The only strong similarity between Pachyderm and Spark is that our versioning of data is somewhat similar to how Spark uses RDD’s and data frames to speed up computation. Spark is a fantastic interface for exploring your data or running queries. In our opinion, Spark is one of the best parts of the Hadoop ecosystem and in the near future, we’ll be offering a connector that lets you use the Spark interface on top of Pachyderm.
What are the major use cases for Pachyderm?¶
Data Lake: A data lake is a place to dump and process gigantic data sets. This is where you send your nightly production database dumps, store all your raw log files and whatever other data you want. You can then process that data using any code you can put in a container. Martin Fowler has a great blog post describing data lakes.
Containerized ETL: ETL (extract, transform, load) is the process of taking raw data and turning it into a useable form for other services to ingest. ETL processes usually involve many steps forming a DAG (Directed Acyclical Graph) — pulling raw data from different sources, teasing out and structuring the useful details, and then pushing those structures into a data warehouse or BI (business intelligence) tool for querying and analysis.
Pachyderm completely manages your ETL DAG by giving you explicit control over the inputs for every stage of your pipeline. We also give you a simple API — just read and write to the local file system inside a container — so it’s easy to push and pull data from a variety of sources.
Automated ML pipelines: Developing machine learning pipelines is always an iterative cycle of experimenting, training/testing, and productionizing. Pachyderm is ideally suited for exactly this type of process.
Data scientists can create jobs to explore and process data. Pachyderm will automatically let you down-sample data or develop analysis locally without having to copy any data around.
Building training/testing data sets is incredibly easy with version-controlled data. Since you have all your historical data at your fingertips, you can simply train a model on data from last week and then test it on this week’s data. Getting training/testing pairs involves zero data copying or moving.
Finally, once your analysis is ready to go, you simply add your job to Pachyderm as a pipeline. Now it’ll automatically run and continue updating as new data comes into the system, letting you seamlessly transition from experimentation to a productionized deployment of your new model.
Is Pachyderm enterprise production ready?¶
Yes! Pachyderm hit v1.2 and is ready for production use! If you need help with your deployment or just want to talk to us about the details, we’d love to hear from you on Slack or email us at email@example.com.
How does Pachyderm handle logging?¶
Kubernetes actually handles all the logging for us. You can use
pachctl get-logs to get logs from your jobs. Kubernetes also comes with it’s own tools
for pushing those logs to whatever other services you use for log aggregation
Does Pachyderm only work with Docker containers?¶
Right now yes, but that’s mostly because Kubernetes doesn’t yet support other runtimes. Pachyderm has no strict dependencies on Docker so we’ll have support for rkt and other container formats soon.
How do I get enterprise support for Pachyderm?¶
If you’re using Pachyderm in production or evaluating it as a potential solution, we’d love to chat with you! firstname.lastname@example.org