Simply delete all entries in a database.

Question

Hello,
how do I simply delete all entries in a database (which must be thread safe, and most probably is)? For instance it is needed, as I'm developing a versioned open source XML/JSON database system, whereas I'm using a BerkeleyDB environment/database as a transaction log per revision (resource/log/version_number/...) for dirty blocks/pages and now want to introduce checkpointing (with currently only a single write-transaction per resource). That is another reading transaction (possibly another thread might read a transaction log, a BerkeleyDB environment/database while another thread, the checkpointer (most probably a deamon thread) commits, that is writes the log periodically or during less workload into the real resource. After the data is commited, the transaction-log must be emptied, but probably a reading transaction still reads from the log and falls back to reading from the real resource if the page is not in the log. That is I can't remove the database, but probably simply have to delete all entries and a simple .commit-file flag which indicates if the data has been written back to the real resource or the checkpointer must be writing it back sometime in the future (if the .commit-file still exists). Do I have to iterate through the database with a cursor (and .getNext())? Or does a dedicated method exist?
kind regards
Johannes

Answer 1

Hi Johannes,
As I think you've already discovered, there is no built-in method for deleting all records of a database that is open. The only similar built-in methods are those for removing or truncating an entire database (removeDatabase and truncateDatabase), and the database must be closed.
If you can't find a way to use removeDatabase or truncateDatabase, then you'll have to iterate through the records and delete them individually. If this is done for a large numbers of records in a single transaction, it will be expensive on a number of fronts, including memory usage for the locks: each record is individually locked.
If you don't need to delete all records in a single transaction (I couldn't completely understand your use case), then you can iterate with a cursor using READ_UNCOMMITTED and delete the records in individual transactions using Database.delete. This avoids using lots of memory for the locks, since only one record is locked at a time.
In either case the cost can be reduced by using DatabaseEntry.setPartial(0, 0, true) for the DatabaseEntry that is passed as the data parameter. You only need the key to delete the record, not the data, and avoiding a fetch of the data is a big cost savings (if the record data is not in cache). This optimization is only in JE 5.0 and above --in JE 4.1 and earlier, this has no advantage because the data is always fetched internally, as part of the deletion operation.
--mark