Failure of expensive test "datamash-valgrind.sh" for git HEAD

[Erik Auerswald]

Hi,

the "expensive" test "datamash-valgrind.sh" fails on my Ubuntu 18.04
GNU/Linux x86-64 laptop with current git HEAD:

```
$ make check-expensive TESTS=tests/datamash-valgrind.sh
[...]
custom-format failed
FAIL: tests/datamash-valgrind.sh
[...]
```

This is the last sub-test in the shell script.  The pertinent log
can be extracted with `sed -n '/--format/,$p' test-suite.log`.  The
test checks if the command `datamash --format "%05000.5000f" sum 1`
with input from file "wide" via standard input leaks memory.

At least one input field of the generated data file "wide" contains
a number that is too large to be represented as an 80 bit extended
precision IEEE 754.  Thus GNU Datamash fails without creating output
in the specified custom format with the error message:

```
datamash: invalid numeric value in line 9 field 1: '123[...]545'
```

Well, there are three such lines:

```
$ for i in $(seq 1 193 2000) $(seq 500 -7 100); do \
>   seq $i | paste -s -d "" - ; \
> done \
> | awk 'length($0)>4932 {
         print "line="NR "fields="NF" length="length($0)
         }'
line=9 fields=1 length=5073
line=10 fields=1 length=5845
line=11 fields=1 length=6617
```

Since datamash fails and exits with exit code 1, valgrind reports
exit code 1 as well, and the test fails.  I do not think this is
intended.

Additionally, valgrind reports reachable memory:

```
==6391== HEAP SUMMARY:
==6391==     in use at exit: 13,400 bytes in 9 blocks
==6391==   total heap usage: 25 allocs, 16 frees, 32,247 bytes allocated
[...]
==6391== LEAK SUMMARY:
==6391==    definitely lost: 0 bytes in 0 blocks
==6391==    indirectly lost: 0 bytes in 0 blocks
==6391==      possibly lost: 0 bytes in 0 blocks
==6391==    still reachable: 13,400 bytes in 9 blocks
==6391==         suppressed: 0 bytes in 0 blocks
==6391==
==6391== For counts of detected and suppressed errors, rerun with: -v
==6391== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)
```

I do not know if the valgrind report indicates a problem or not.

I think that we should adjust this test.  We need to either make
datamash exit successfully, e.g., by adjusting the input, or to ignore
the datamash failure, but still detect a valgrind failure.  The former
might be possible by removing the problematic lines from the input via
sed.  The latter might be possible via valgrind's --error-exitcode=N
option.

I have successfully tested the former using the following patch:

```
diff --git a/tests/datamash-valgrind.sh b/tests/datamash-valgrind.sh
index 6bb74af..43e9fd9 100755
--- a/tests/datamash-valgrind.sh
+++ b/tests/datamash-valgrind.sh
@@ -188,7 +188,7 @@ cmp wide wide_orig ||
     fail=1 ; }

 ## Test large output formats
-cat wide | valgrind --track-origins=yes  --leak-check=full \
+sed 9,11d wide | valgrind --track-origins=yes  --leak-check=full \
                   --show-reachable=yes  --error-exitcode=1 \
                   datamash --format "%05000.5000f" sum 1 > /dev/null ||
   { warn_ "custom-format failed" ; fail=1 ; }
```

The valgrind report for this contains:

```
[...]
==10177==    still reachable: 14 bytes in 1 blocks
[...]
==10177== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)
```

Again I am not sure about the meaning of the report.

What do you think?

Cheers,
Erik