Project

General

Profile

Bug #21547

Updated by Peter Amstutz 10 months ago

Certain database errors represent transient errors.    We should tell the client to retry the request by returning a 500 internal server error instead of 422 (which is the default behavior). 

 #<ActiveRecord::Deadlocked: PG::TRDeadlockDetected: ERROR: deadlock detected> 

 Rationale: The observed deadlocks in Arvados are conflicts between two statements (a lock ordering issue), so unwinding and retrying is a reasonable solution 

 #<ActiveRecord::StatementInvalid: PG::UnableToSend> 

 Rationale: It seems this gets thrown when the API server can't connect to the database. 

 Here's the list of postgres errors known to the PG gem: 

 https://github.com/ged/ruby-pg/blob/daec80f91b9519509ca1694a231f11a75cb43f7f/ext/errorcodes.def#L598 

 https://github.com/ged/ruby-pg/blob/daec80f91b9519509ca1694a231f11a75cb43f7f/ext/pg_errors.c#L88 

 Some other possible Exceptions to retry: 

 ConnectionBad 
 ConnectionException 
 ConnectionDoesNotExist 
 ConnectionFailure 
 TooManyConnections 
 CannotConnectNow 
 IdleSessionTimeout 
 ObjectInUse 
 LockNotAvailable 
 AdminShutdown 
 CrashShutdown 

 (There's a lot of connection related errors and I don't know the difference between them, but I included them all because it seems like those are very likely to be errors that occur through no fault of the client). 

Back