Long startup cost for simple session #3713
Last updated: 2015-06-22 22:32:19 +0200
Date: 2015-04-22 21:41:14 +0200
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.10; rv:37.0) Gecko/20100101 Firefox/37.0
Consider the TPCH sf1 script below, where sf1.load contains the relevant COPYs.
Consistently, we need to wait for 9 seconds on a trivial load:
compared to the rest of the load:
The subsequent alter script runs fine, also for the first instruction
If I place a sleep of 15 seconds after tpch_schema.sql,
If I explicitly stop the server and let it be restarted
Seems that the tpch_schema.sql transaction does not propagate the catalog updates immediately, giving this unexpected delay. (who triggers the wait?)
Steps to Reproduce:
monetdb stop sf1
Date: 2015-04-22 22:17:38 +0200
I can reproduce the problem.
However, if I do the load UNLOCKED (i.e. not LOCKED), the problem disappears.
From looking at strace output I get the distinct impression that it has something to do with committing data. When using LOCKED, after the schema initialization I see that the first data file is loaded, but before that query returns, there are a lot of BATs that get written and fdatasync()ed. All in all that takes a long time. Without LOCKED I don't see the long list of BATs getting synchronized to disk.
Here are the timings of my CREATE script immediately followed by loading with LOCKED:
And here without LOCKED:
Date: 2015-04-24 13:41:11 +0200
changes to the logger caused extra delay. BUNfnd(.., neg-cand list) may
Date: 2015-04-24 15:21:34 +0200
we use a hash(again) instead of nestedloop (the deleted list is still in efficient)
The text was updated successfully, but these errors were encountered: