Databricks connect 14.3.2 SparkConnectGrpcException Not found any cached local relation withthe hash

[ahsan_aj]

Hi All,

I am using Databricks Connect 14.3.2 with Databricks Runtime 14.3 LTS to execute the code below. The CSV file is only 7MB, the code runs without issues on Databricks Runtime 15+ clusters but consistently produces the error message shown below when using 14.3 LTS. Please advise.

SparkConnectGrpcException: (org.apache.spark.sql.connect.common.InvalidPlanInput) Not found any cached local relation with the hash: hash_guid in the session with sessionUUID session_guid.

import pandas as pd
from databricks.connect import DatabricksSession

cluster_id = '*****'
user_token = "*******"
host = 'https://***********.azuredatabricks.net/'

sp = DatabricksSession.builder.remote(host = host,cluster_id = cluster_id,token = user_token).getOrCreate()
df = sp.createDataFrame(pd.read_csv("C:\\temp\\data.csv"))
df.show(5)

[ahsan_aj]

Unfortunately I don't have this information. I only raised it for 14.3 LTS since my databricks connect version is the same (14.3.1).

[ADuma]

The error is now occuring for Cluster with Version 15.4. for me. Did the fix get released yet?

[ahsan_aj]

I don't think it has been released yet, and now I’m facing this issue on both 14.3 LTS and 15.4 LTS :(.

FYI @Retired_mod 

[asia_sowa]

I have the same issue with 13.3 LTS version

[LukeEs]

I was informed that the fix was released. According to our tests, however, it did not fix anything. Instead, 15.4. LTS is now broken, too. This topic is getting urgent for us, now.

[felix_]

Any updates on this? Facing the same issue with 15.4 LTS now as well..

[ahsan_aj]

Microsoft support just mentioned that fix has been deployed by Databricks, but the issue continues to persist for me on both 14.3 LTS and 15.4 LTS.

[CarlDaniel]

Now same issue with version 15.4 LTS. Does the fix for 14.3 LTS work? Thanks!

[MichalMazurek]

I still have the issue, but I noticed I do not have it on Linux. Support told me to use this line:

spark.conf.set("spark.sql.session.localRelationCacheThreshold", 64 * 1024 * 1024)

With that, it worked on windows, also this gives a hint what should be the batch size.

[ahsan_aj]

I have a troubleshooting session scheduled with Databricks today regarding this issue and will keep everyone updated on the progress.

[ahsan_aj · Accepted solution]

As a workaround, please try the following Spark configuration, which seems to have resolved the issue for me on both 14.3 LTS and 15.4 LTS.

spark.conf.set("spark.sql.session.localRelationCacheThreshold", 64 * 1024 * 1024)

[ahsan_aj]

Databricks confirmed the same workaround while they work on a permanent fix.