On Wed, Mar 15, 2017 at 9:17 AM, Enrico Olivelli <eolive...@gmail.com> wrote:
> Hi BookKeepers, > I'm trying to understand some basic aspect of BookKeeper.... > > I have a strange case, and I would like to have some opinion from any > "expert". > > I am using BookKeeper as a transaction log for a database, this is an > example scenario: > > 1) the user inserts a record in a table > the "fact" is logged to the Write-Ahead-log (Bookkeeper ledger), and the > systems blocks until the asyncAddEntry completes (the callback is called > with rc=0) > > now I (the 'writer') assume that the entry is written to durable storage > and has been synched to disk > then the DB applies the change to the local memory as every application > with a WAL does > > 2) The writer 2 other entries for which the asyncAddEntry completes (the > callback is called with rc=0). I block until the callback as been called > > 2) for tests I'm using an ensemble of 1 bookie, I "stop" the only bookie > the bookie stops "gracefully" (no strange errors) > > 3) The writer issues other asyncAddEntry, which fail due to > "notEnoughBookiesException" and then "closes" the LedgerHandle, inside the > "callback" cade > > 4) I restart the bookie (no strange errors on logs) > > 5) I restart my "writer", opening the ledger with "openLedger" (with > recovery and fencing...) in order to perform recovery from the WAL > > 6) now the LastAddConfirmed is 0 ! so my first entry is lost ! > > Am I missing something ? > The LAC has not been advanced even with 3 successful writes to the ledger > Are you using readLastAddConfirmed at step 5)? 0 is expected when you using #readLastAddConfirmed, since the call is retrieving LAC from bookies. At step 3), it already closed the ledger and update the metadata's last entry id with 2. so for tailing read, you might need to call #readLastAddConfirmed and then get the last add confirmed using #getLastAddConfirmed. > > From the "writer" point-of-view the fact that the callback as been called > is not a guarantee that the entry as been written to durable storage ? > > I understand that of a "tailing" reader (a ledger opened with > openLedgerNoRecovery) it is very important not to be able to "read" data > before the LAC, but for the "writer" it is important that it can count on > the fact the entry will be recovered with success > > > Thanks > -- Enrico > >
