I'd like to use Deep Learning on Spark within AWS EMR.Is there a best practice or 'recommended' DL framework to run on Spark? It looks like Databricks' spark-deep-learning has been replaced by Horovod—should this the first option to consider? If th...
I am not aware of any special requirement for this migration, my suggestion to you is to try it on a small scale (one notebook) and observe the results showing in the tracker server, if everything looks OK, then migrate the rest.
If you have configured your Structured Streaming query to use RocksDB as the state store, you can now get better visibility into the performance of RocksDB, with detailed metrics on get/put latencies, compaction latencies, cache hits, and so on. Thes...
Advantage of using Photon EngineThe following summarizes the advantages of Photon:Supports SQL and equivalent DataFrame operations against Delta and Parquet tables.Expected to accelerate queries that process a significant amount of data (100GB+) and ...
check if your workspace has the IP access list feature enabled, call the get feature status API (GET /workspace-conf). Pass keys=enableIpAccessLists as arguments to the request.In the response, the enableIpAccessListsthe field specifies either true o...
Wondering about best practices for how to handle collaboration between multiple ML practitioners working on a single experiment. Do we have to share the same notebook between people or is it possible to have individual notebooks going but still work ...
Yes, multiple users could work on individual notebooks and still use the same experiment via mlflow.set_experiment(). You could also assign different permission levels to experiments from a governance point of view
The default location or MLflow artifacts is on dbfs, but I would like to save my models to an alternative location. Is this supported, and if it is how can I accomplish it?
You could mount an s3 bucket in the workspace and save your model using the mounts DBFS path For e.gmodelpath = "/dbfs/my-s3-bucket/model-%f-%f" % (alpha, l1_ratio) mlflow.sklearn.save_model(lr, modelpath)
Not yet, but stay-tuned it's being cooked in the kitchen
Data is stored in the control plane. Metadata (eg feature table descriptions, column types, etc) is stored in the control plane. The location where the Delta table is stored is determined by the database location. The customer could call CREATE DATA...
Is it possible to rollback changes made to a cluster? The problem I'm trying to solve is to recover from an accidental change made by a user on a cluster that affects interactive and job runs. Cluster policies help, but the policy still provides the ...
You could look at automating cluster creation steps and implementing this with an infra-as-code solution like the databricks terraform provider which allows rollback
Do we have general guidance around how other customers manage Dev and Prod environments in Databricks? Is it recommended to have separate workspaces for them? What are the pros and cons of using the same workspace with folder or repo level isolation?
I'm trying to run Delta Lake MergeMERGE INTO source USING updates ON source.d = updates.sessionId WHEN MATCHED THEN UPDATE * WHEN NOT MATCHED THEN INSERT *I'm getting an SQL errorParseException: mismatched input 'MERGE' expecting {'(', 'SELECT', 'FR...
The merge SQL support is added in Delta Lake 0.7.0. You also need to upgrade your Apache Spark to 3.0.0 and enable the integration with Apache Spark DataSourceV2 and C
There's a lot of different ML formats out there and I am confused about how they should be fitting together. How should I be thinking about MLflow and MLeap working together?
I am trying to set up a demo with a really simple spark ML model and i see this error repeated over and over in the logs in the serving UI:/databricks/chauffeur/model-runner/lib/python3.6/site-packages/urllib3/connectionpool.py:1020: InsecureRequestW...
Not sure how the containers for each model version work on the endpoints, but looks like Model serving endpoints use a 7.x runtime. So those would be Spark 3.0, not Spark 3.1
I can see an example on how to call the vacuum function for a Delta lake in python here. how to use the same in python %sql VACUUM delta.`dbfs:/mnt/<myfolder>` DRY RUN
The dry run for non-SQL code is not yet available in Delta version 0.8. I see there is a bug that is opened with delta opensource in git . hope it get resolved soon
