Paul Rogers created DRILL-6168:
----------------------------------
Summary: Table functions do not "inherit" default configuration
Key: DRILL-6168
URL: https://issues.apache.org/jira/browse/DRILL-6168
Project: Apache Drill
Issue Type: Bug
Affects Versions: 1.12.0
Reporter: Paul Rogers
See DRILL-6167 that describes an attempt to use a table function with a regex
format plugin.
Consider the plugin configuration:
{code}
RegexFormatConfig sampleConfig = new RegexFormatConfig();
sampleConfig.extension = "log1";
sampleConfig.regex = DATE_ONLY_PATTERN;
sampleConfig.fields = Lists.newArrayList("year", "month", "day");
{code}
(This plugin is defined in code in a test rather than the usual JSON in the Web
console.)
Run a test with the above. Things work fine.
Now, try the plugin config with a table function as described in DRILL-6167:
{code}
String sql = "SELECT * FROM table(cp.`regex/simple.log2`\n" +
"(type => 'regex', regex =>
'(\\\\d\\\\d\\\\d\\\\d)-(\\\\d\\\\d)-(\\\\d\\\\d) .*'))";
client.queryBuilder().sql(sql).printCsv();
{code}
Because we are using a file with suffix "log2", the query will match the format
plugin config defined above. A query without the table function does, in fact,
work using the defined config. But, with a table function, we get this warning
from our regex code:
{noformat}
13307 WARN [257590e1-e846-9d82-61d4-e246a4925ac3:frag:0:0]
[org.apache.drill.exec.store.easy.regex.RegexRecordReader] - Column list has
fewer
names than the pattern has groups, filling extras with Column$n.
{noformat}
(The warning is in the custom plugin, not Drill.) This is the plugin saying,
"hey! you didn't provide column names!". But, in the format definition, we did
provide names. If we run the query without a table function, we do see those
names used.
Result:
{noformat}
3 row(s):
Column$0<VARCHAR(OPTIONAL)>,Column$1<VARCHAR(OPTIONAL)>,Column$2<VARCHAR(OPTIONAL)>
2017,12,17
2017,12,18
2017,12,19
Total rows returned : 3. Returned in 9072ms.
{noformat}
Yes, indeed, the table function discarded the defined format config values,
filling in blanks, including for the column names.
The expected behavior is that all properties defined in the config should
remain unchanged _except_ for those in the table function. Why? In order to
know which format plugin to use, the code has to map from the suffix (".log2"
here) to a format plugin _config_. (The config is the only thing that specifies
a suffix.) Since we mapped to a config (not the unconfigured plugin), we'd
expect the config properties to be used.
It is highly surprising that all we get to use is the suffix, but all other
attributes are ignored. This seems very much in the "bug" category and not at
all in the "feature" category.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
