Note in the description, I had this issue ONLY when exceeding the max_resultset_size parameter. Not when exceeding the maxrows parameter.

MaxScale configuration:

[MaxRows]
type=filter
module=maxrows
max_resultset_rows=10
max_resultset_size=900000
debug=3

[Read-Only-Service]
type=service
router=readconnroute
router_options=running
servers=server1,server2,server3
user=massi
passwd=massi
filters=MaxRows

MySQL client to maxscale:

mysql -h 127.0.0.1 -P 8806 -umassi -pmassi

MySQL [(none)]> select repeat('a', 1000000);
Empty set (0.01 sec)

Messages in the log file:
2017-03-31 08:58:54 info : [readconnroute] Routed [COM_QUERY] to 'server1': select repeat('a', 1000000)
2017-03-31 08:58:54 notice : [maxrows] RESULTSET
2017-03-31 08:58:54 notice : [maxrows] Current size 1000068B of resultset, at least as much as maximum allowed size 878KiB. Not returning data.
2017-03-31 08:58:54 notice : [maxrows] OK or EOF packet seen: the resultset has 1 rows. [Discarded]

I see the empty result set comes immediately.

Would you mind adding MaxScale configuration and a SQL example that takes that long time, 23 seconds?

I have a table with 12 columns and 66k rows. When I execute select * from table with the filters=MaxRows line commented out under the [Splitter Service] the entire query is returned in about 5s.
When I uncomment the filters=MaxRows line under [Splitter Service], an empty result set is returned in about 23s.
Maxscale is running on an m4.xlarge server, 4 cpu cores. I have noted that during the 23s I am waiting for the empty result set to return, 1 of the 4 cores is pegged at 100%. Something is cpu bound in the max_resultset_size parameter.
The max_resultset_rows parameter was commented out in both tests.
I am able to consistently reproduce the 5s and 23s execution times.
maxscale.cnf

Hi Darin, thanks for the update.

As long as you are able to reproduce the 23s execution time I would like ask you to add in maxscale.cfnf

[maxscale]
threads=4
syslog=1
maxlog=1
log_to_shm=1
log_warning=1
log_notice=1

[MaxRows]
type=filter
module=maxrows
debug=3

then you could connect to MaxScale using a mysql/mariadb client:

MySQL [(none)]> select repeat('a', 50000) from test.t4 limit 10;
Empty set (0.01 sec)

Please report the content of maxscale.log: this is as an example from my setup, using that configuration:

_2017-04-02 18:08:26 info : [readconnroute] Routed [COM_QUERY] to 'server1': select repeat('a', 50000) from test.t4 limit 10
2017-04-02 18:08:26 notice : [maxrows] RESULTSET
2017-04-02 18:08:26 notice : [maxrows] Current size 100072B of resultset, at least as much as maximum allowed size 64KiB. Not returning data.
2017-04-02 18:08:26 notice : [maxrows] OK or EOF packet seen: the resultset has 10 rows. [Discarded]
_

This way the MaxScale log could tell us something more.

Thanks.

Massimiliano

maxscale.cnf maxscale.log

Log and cnf attached.

2017-04-06 18:36:22 notice : [maxrows] Maxrows filter is sending data.
2017-04-06 18:36:40 notice : [maxrows] OK or EOF packet seen: the resultset has 66548 rows. [Discarded]
You can see from the timestamps, 18 seconds elapsed before the EOF notice. I commented out the filter and got the full resultset in ~6s.

I'm also surprised the log says the result set exceeded rows when I have the rows limited commented out.

Fixed usage of gwbuf_length: partial buffer only, not the whole buffer.

This lowers CU usage when receiving 30.000 and more rows

+ csdata->res.length += gwbuf_length(data);
....

• size_t buflen = gwbuf_length(csdata->res.data);
  + size_t buflen = csdata->res.length;

Code pushed to 2.1 branch:

MXS-1212: use res.length for buffer length instead of computing the size of the whole buffer chain

This reduces CPU usage and decreases the response time for result sets
with many rows.

https://github.com/mariadb-corporation/MaxScale/commit/f4a5134a83a79dd45a07cb75edfb1777773c049f