Paul Rogers created DRILL-6075:
----------------------------------
Summary: Drill-provided UDF helper functions may prevent scalar
replacement
Key: DRILL-6075
URL: https://issues.apache.org/jira/browse/DRILL-6075
Project: Apache Drill
Issue Type: Bug
Affects Versions: 1.10.0
Reporter: Paul Rogers
Priority: Minor
Users can write UDFs which use "holders" to pass data into and out of
functions. UDFs are templates, written in a Drill-defined DSL, that Drill uses
to copy UDF code into the generated code (often in the Project operator.)
Our
[documentation|http://drill.apache.org/docs/tutorial-develop-a-simple-function/]
has vague warnings about not referencing methods in the holders because doing
so prevents scalar replacement.
But, at the same time, Drill provides the {{StringFunctionHelpers}} class that
contains, among others, these two methods:
{code}
public static String getStringFromVarCharHolder(VarCharHolder varCharHolder);
public static String getStringFromVarCharHolder(NullableVarCharHolder
varCharHolder);
{code}
The example shown in the documentation calls one of the above methods.
We're sending mixed signals (or are confused) here. By passing a holder to a
method, we prevent any reasonable scalar replacement algorithm from working.
So, if we believe that not replacing holders is slow, we should not encourage
use of the above methods.
Instead, the rules for holders should be:
* Call no methods on the holders.
* Do not pass holders to other functions.
A number of Drill's own functions make this mistake:
{{SqlToDate}}
{{SqlToTime}}
{{SqlToTimestamp}}
{{DirectoryExplorers.[IMax*, IMin*, Max*, Min*]}}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
