[
https://issues.apache.org/jira/browse/CASSANDRA-15194?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16899164#comment-16899164
]
Jon Haddad commented on CASSANDRA-15194:
----------------------------------------
Just took a look at the patch and noticed a few things.
I ran a simple test using tlp-stress, a basic key value run:
{noformat}
bin/tlp-stress run KeyValue -d 1h --rate 1000 -r .5
{noformat}
When I checked the virtual tables I'm seeing a p99 of 73 seconds:
{noformat}
cqlsh:system_views> select * from local_read_latency where keyspace_name =
'tlp_stress' allow filtering;
keyspace_name | table_name | 99th_ms | count | max_ms | median_ms |
per_second
---------------+-------------+---------+-------+------------+-----------+------------
tlp_stress | keyvalue | 73457 | 86523 | 1.3581e+06 | 11864 |
205.219
tlp_stress | sensor_data | 0 | 0 | 0 | 0 |
0
{noformat}
However, I don't see anything close to that in tlp-stress:
{noformat}
Writes Reads
Errors
Count Latency (p99) 1min (req/s) | Count Latency (p99) 1min (req/s) |
Count 1min (errors/s)
76557 0.38 501.04 | 76620 0.36 501.75 |
0 0
78058 0.38 500.41 | 78119 0.36 502.15 |
0 0
79503 0.37 500.41 | 79671 0.36 502.15 |
0 0
80999 0.37 499.96 | 81174 0.37 502.4 |
0 0
82553 0.37 500.4 | 82623 0.37 501.77 |
0 0
84085 0.36 500.4 | 84092 0.38 501.77 |
0 0
85544 0.37 500.49 | 85633 0.38 501.5 |
0 0
{noformat}
Nodetool tablehistograms agrees with tlp_stress that my laptop isn't perfoming
that poorly:
{noformat}
$ nodetool tablehistograms tlp_stress keyvalue
tlp_stress/keyvalue histograms
Percentile Read Latency Write Latency SSTables Partition
Size Cell Count
(micros) (micros)
(bytes)
50% 11.86 8.24 0.00
215 1
75% 17.08 9.89 0.00
215 1
95% 42.51 29.52 0.00
258 1
98% 61.21 35.43 0.00
258 1
99% 73.46 35.43 0.00
258 1
Min 3.97 2.76 0.00
125 0
Max 152.32 73.46 1.00
258 1
{noformat}
The same issue pops up for other tables as well:
{noformat}
cqlsh:system_views> select * from local_scan_latency where keyspace_name =
'tlp_stress' allow filtering;
keyspace_name | table_name | 99th_ms | count | max_ms | median_ms |
per_second
---------------+-------------+------------+-------+------------+------------+------------
tlp_stress | keyvalue | 1.1318e+06 | 16 | 1.1318e+06 | 5.4579e+05 |
0.05
tlp_stress | sensor_data | 0 | 0 | 0 | 0 |
0
{noformat}
I think starting at line 205 of TableMetricTables.java you'd want this:
{noformat}
add(result, MEDIAN + suffix, snapshot.getMedian() / NS_TO_MS);
add(result, P99 + suffix, snapshot.get99thPercentile() / NS_TO_MS);
add(result, MAX + suffix, (double) snapshot.getMax() / NS_TO_MS);
{noformat}
When I apply that change and rerun things, I get output that makes a lot more
sense:
{noformat}
cqlsh:system_views> select * from local_read_latency where keyspace_name =
'tlp_stress' allow filtering;
keyspace_name | table_name | 99th_ms | count | max_ms | median_ms | per_second
---------------+-------------+---------+-------+--------+-----------+------------
tlp_stress | keyvalue | 0.786 | 17660 | 7.008 | 0.03 | 53.331
tlp_stress | sensor_data | 0 | 0 | 0 | 0 | 0
{noformat}
Other than that, I don't have any issues. We can merge once that's addressed.
We should create a follow up JIRA to document all the virtual tables.
> Improve readability of Table metrics Virtual tables units
> ---------------------------------------------------------
>
> Key: CASSANDRA-15194
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15194
> Project: Cassandra
> Issue Type: Bug
> Components: Feature/Virtual Tables
> Reporter: Jon Haddad
> Assignee: Chris Lohfink
> Priority: Normal
> Fix For: 4.0
>
>
> I just noticed this strange output in the coordinator_reads output::
> {code}
> cqlsh:system_views> select * from coordinator_reads ;
> count | keyspace_name | table_name | 99th | max |
> median | per_second
> -------+--------------------+--------------------------------+------+-----+--------+------------
> 7573 | tlp_stress | keyvalue | 0 | 0 |
> 0 | 2.2375e-16
> 6076 | tlp_stress | random_access | 0 | 0 |
> 0 | 7.4126e-12
> 390 | tlp_stress | sensor_data_udt | 0 | 0 |
> 0 | 1.7721e-64
> 30 | system | local | 0 | 0 |
> 0 | 0.006406
> 11 | system_schema | columns | 0 | 0 |
> 0 | 1.1192e-16
> 11 | system_schema | indexes | 0 | 0 |
> 0 | 1.1192e-16
> 11 | system_schema | tables | 0 | 0 |
> 0 | 1.1192e-16
> 11 | system_schema | views | 0 | 0 |
> 0 | 1.1192e-16
> {code}
> cc [~cnlwsu]
> btw I realize the output is technically correct, but it's not very readable.
> For practical purposes this should just say 0.
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
