[erlang-questions] mnesia - secondary indexes vs separate tables

[Ignas Vyšniauskas]

05/28/2013 05:19 PM, Ulf Wiger wrote:
> On 28 May 2013, at 15:14, Motiejus Jakštys wrote:
>
>> Some folks did some benchmarks with mnesia with a secondary index,
>> and write performance was terrible (~40 times slowdown for
>> inserts?). But I could not find it in the archives. Who was it?
>
> I believe it was Scott Fritchie, but it was for the special case
> where an inordinate number of secondary keys were of the same value
> (e.g. indexing a boolean attribute). Since index tables are 'bag'
> tables, this can indeed result in terrible performance.

Indeed it was Scott, here:
http://erlang.2086793.n4.nabble.com/Mnesia-and-additional-indexes-a-cautionary-tale-td2088434.html

And I have done some benchmarks which confirmed crappy insert
performance with secondary indices, but it was not very thorough,
because I bumped into a more severe problem, see below.

On 05/28/2013 02:26 PM, John Doe wrote:
> The table(s) will be disc_only_copy, as RAM matters.

Last time I checked (R14 I think) mnesia never stored the secondary
indices, so it would recreate them every time you restarted the node,
and given that you have an order of 10^6 record this will have either a
significant or a very significant impact on your start time, so you
might want check if that's still the case and decide whether it's a
problem for you.

--
Ignas