It seems that Impala indeed starts without HDFS when the flag is reset as suggested. I'm also observing some errors in the logs, due to the inability to find/connect to HDFS:

Could not read the root directory at hdfs://
Error was: Call From localhost/ZZ.ZZZ.ZZ.ZZZ to failed on connection exception: Connection refused; For more details see:

This is likely due to the fact that HDFS is still mentioned in the configs (specifically, core-site.xml still refers to as to hdfs:// Should this be modified somehow to refer to the local file system or removed altogether?


On Thu, Apr 27, 2017 at 8:53 AM, Aleksei Maželis <> wrote:
Ok, great, thanks for elaborating! I will give it a try with the abort_on_config_error flag reset! 

As to the Hive dependency, I will stay tuned for the future updates then.


On Wed, Apr 26, 2017 at 6:31 PM, Alexander Behm <> wrote:
You should be able to bring Impala up without HDFS by passing the "--abort_on_config_error=false" startup flag.

Just to clarify:
You still need the Hive Metastore with Kudu because table and column statistics are stored in the Metastore; our plan is to eventually remove this dependency

On Wed, Apr 26, 2017 at 1:28 AM, Aleksei Maželis <> wrote:
Ok, I see. I remember trying to run Impala-Kudu without HDFS, and remember that Impala failed to start. But perhaps there is a way to make Impala work without HDFS after re-configuring it properly; any hints on which configurations to change would be very helpful!  


On Tue, Apr 25, 2017 at 7:36 PM, Jim Apple <> wrote:
I believe the Hive Metastore is needed even when Kudu is the storage engine. I don't know if the HDFS namenode and datanodes are needed.

On Mon, Apr 24, 2017 at 10:28 PM, Aleksei Maželis <> wrote:

I am looking for the minimal set of dependencies that an Impala-Kudu setup will have. When reading how-tos and checking available dockerfiles, the list of items that Impala depends on seems to include at least:
- hadoop-hdfs-namenode
- hadoop-hdfs-datanode
- hive-metastore.

However, if I understand correctly, when Kudu is used along with Impala, these aren't necessarily used. So, the question is if the above dependencies of HDFS and Hive can be avoided, and if so, whether there are side effects this would bring. And of course, if I am missing some other mandatory dependencies, hints on those would be more then welcome!

With best regards,