[
https://issues.apache.org/jira/browse/MADLIB-1501?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17558172#comment-17558172
]
Orhan Kislal commented on MADLIB-1501:
--------------------------------------
Hi [~Blairruc-pku],
Unfortunately, PostgreSQL has a hard limit of 1GB on the field size:
[https://www.postgresql.org/docs/current/limits.html.]
Since MADlib models have to be stored in a field to iterate over them we cannot
work around this issue.
Please note that MADlib is designed to work with both PostgreSQL and Greenplum,
so we have to consider building partial models and merging them for the
multiple segments of GPDB. We will continue exploring ideas to overcome this
problem, but for now, we must abide by this limitation.
> Can not train model larger than 1GB.
> ------------------------------------
>
> Key: MADLIB-1501
> URL: https://issues.apache.org/jira/browse/MADLIB-1501
> Project: Apache MADlib
> Issue Type: Bug
> Components: Deep Learning
> Reporter: Xinyi Zhang
> Priority: Major
> Fix For: v1.19.0
>
>
> When I want to train a model whose size is large than 1GB on Greenplum, I get
> the error below:
> CONTEXT: PL/Python function "madlib_keras_fit"
> ERROR: spiexceptions.InternalError: invalid memory alloc request size
> 1100478264 (plpy_elog.c:121) .
>
> But If I use a smaller model, it can run successfully.
> It seems that "SELECT \{schema_madlib}.fit_step()" can not execute when the
> model is larger than 1GB.
> I set my shared_buffers to 32GB, and the instance has 290G memory available.
> So, something wrong might happen to the memory allocation in Madlib.
> I did not find any parameters to solve the problem. But since the large model
> is quite common, I think there should be a solution for training models
> larger than 1GB in Madlib.
>
>
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
