Re: A question about node restarting after crashed

[Mikael Ronström]

The following things happen at a node restart (not complete, only major steps):
1) The living node specifies the tables and fragments to restore
2) A LCP is restored
3) The REDO log is applied up to a certain group commit point (GCI) that is available in the starting node.
4) The living node scans row-by-row (with a read-lock on the row for a short time) and those rows that have group commit points (GCIs) that are newer than
the restored one will be sent over to the starting node (this includes also rows that have been deleted). This is described in some detail in a VLDB paper from 2005. In addition the new node is part of all transactions to ensure that the
rows stays up-to-date.
5) When the synchronisation is completed a new LCP is performed.
6) Node restart is done and both nodes can be used for both read and write.