Re: Recovery bugs, classified by me

[Philip Stoev]

>> So, this remains a valid bug for me. I do intend to test recovery
>> systematically with kill -9 immediately after server startup, so a
>> decision
>> and a solution must be implemented for that one. Maybe the solution is
>> to do
>> extra checkpoints after creating the system tables and waiting for the
>> gophers to write everything to disk.
>
> And what you do if you kill before checkpoint has run?

It appears to me that the current behavior is as follows:

1. Falcon starts up, system tables are created in memory
2. Server becomes available for connections
3. Queries start arriving
4. A scheduled checkpoint arrives, the gophers write the system tables to 
disk, etc.

If there is a crash in Step #3, you can not use a workaround "delete 
tablespaces and start from scratch", because you would loose the 
transactions that were issued by the users.

So, instead, maybe this will work:

1. Falcon starts up
2. System tables are created and flushed to disk, force two checkpoints, 
waits for gophers to complete, whatever is needed
3. Server becomes available for connections
4. Queries start arriving

This way, for crashes in Step #2, the workaround can be "delete tablespaces 
and start from scratch". Crashes in Step #4 should recover properly without 
waivers.

Philip Stoev