Queries slow down hundreds times after overwriting points

[kub00n]

Bug report

System info:
InfluxDB 0.13, InfluxDB 0.12.2
Ubuntu 14.04.1

Steps to reproduce:

1. Add some measurements ( 5M points, timerange: 5000 s )

...
test_series,tag=tag1 value=2 1462060800001
test_series,tag=tag2 value=3 1462060800002
test_series,tag=tag0 value=4 1462060800003
...

1. Run a query
2. Insert exactly the same points again (overwriting old ones)
3. Run a query

Expected behavior:
Both query return the same results at a similar time

Actual behavior:
Results are the same, but second query (after overwrite) is almost 500 times slower than the first query. Repeating query after 10s works a bit faster but still far from the first query.

Additional info:
I wrote simple python script to reproduce it : influx_overwrite_bug.py :

==CREATING DATABASE==
==INSERTING SERIES==
* inserting 5000000 points: from 1462060800000ms - to 1462065800000ms
==QUERY==
* query: select * from test_series LIMIT 1
* result: {u'tag': u'tag0', u'value': 1, u'time': 1462060800000}
* duration: 0.0093s
==INSERTING SERIES==
* inserting 5000000 points: from 1462060800000ms - to 1462065800000ms
==QUERY==
* query: select * from test_series LIMIT 1
* result: {u'tag': u'tag0', u'value': 1, u'time': 1462060800000}
* duration: 4.4923s
==SLEEP 10s==
==QUERY==
* query: select * from test_series LIMIT 1
* result: {u'tag': u'tag0', u'value': 1, u'time': 1462060800000}
* duration: 3.1518s
==REMOVING DATABASE==

[jwilder]

@kub00n Thanks for providing the python script. That made it really easy to reproduce and track down the issue. The perf issue is fixed in master, but your script highlighted a correctness issue with deduplicating overwritten points. See #6668 which will fix that issue.