view doc/implementation.txt @ 6565:2c2dbfc332ba

Try to handle multiple connections better. The session database is a hot spot. When multiple requests (e.g. 20) come in at the same time session database contention can get great. The original code didn't retry session database access when the open failed. This resulted in errors at the client. The second pass delayed 0.01 seconds and retried. It was better but we still had multiple second stalls. I think the first request got in, everybody else backed up and then retried at the same time. Again they stepped on each other. With logging I would see many counters go all the way to low single digits or to -1 indicating falure. This pass uses randomint to generate delays from 0-.125 seconds in 5ms increments. This performs better in testing. I rarely saw a counter less than 13 (2 failed retries). Current logging starts after 6 failures and counts down until success or failure.
author John Rouillard <rouilj@ieee.org>
date Thu, 16 Dec 2021 20:02:00 -0500
parents 33a1f03b9de0
children 9ca128103a3a
line wrap: on
line source

====================
Implementation notes
====================

[see also the roundup package docstring]

There have been some modifications to the spec. I've marked these in the
source with 'XXX' comments when I remember to.

In short:
 Class.find() - may match multiple properties, uses keyword args.

 Class.filter() - isn't in the spec and it's very useful to have at the
    Class level.

 CGI interface index view specifier layout part - lose the '+' from the
    sorting arguments (it's a reserved URL character ;). Just made no
    prefix mean ascending and '-' prefix descending.

 ItemClass - renamed to IssueClass to better match it only having one
    hypderdb class "issue". Allowing > 1 hyperdb class breaks the
    "superseder" multilink (since it can only link to one thing, and
    we'd want bugs to link to support and vice-versa).

 template - the call="link()" is handled by special-case mechanisms in
    my top-level CGI handler. In a nutshell, the handler looks for a
    method on itself called 'index%s' or 'item%s' where %s is a class.
    Most items pass on to the templating mechanism, but the file class
    _always_ does downloading. It'll probably stay this way too...

 template - call="link(property)" may be used to link "the current item"
    (from an index) - the link text is the property specified.

 template - added functions that I found very useful: List, History and
    Submit.

 template - items must specify the message lists, history, etc. Having
    them by default was sometimes not wanted.

 template - index view determines its default columns from the
    template's ``tal:condition="request/show/<property>"`` directives.

 template - menu() and field() look awfully similar now .... ;)

 roundup_admin.py - the command-line tool has a lot more commands at its
    disposal

-----------------

Back to `Table of Contents`_

.. _`Table of Contents`: index.html


Roundup Issue Tracker: http://roundup-tracker.org/