This is a continuation of my previous post: SQLite Adapter for NexJ Express – Overcoming the limitations of SQLite part 1.
I made mention of the error codes
SQLITE_BUSY the last time I blogged. These errors show up when new connections try to either read or write to an SQLite database instance that already has a connection with an active write operation. Question 5 from the SQLite FAQ provides a quick description of the overall problem. The gist of it is that when interacting with an SQLite database using multiple concurrent connections, it only takes one writer connection to lock the database file and lock out all the rest. NexJ Express makes use of threading on the persistence layer and as such ran into this issue.
For several weeks we tried various solutions ranging from setting statement query time outs, creating an sqlite3_busy_handler to creating a singleton connection for all threads. The final solution was found after reading about transactions in chapter 5 of “The Definitive Guide to SQLite, Second Edition”. The chapter explains the transaction life cycle and the transition between the different lock states. Write operations cycle through the following states
UNLOCKED -> PENDING (1) -> SHARED -> RESERVED -> PENDING (2) -> EXCLUSIVE -> COMMIT
Of note are PENDING states; PENDING locks are gateway locks.
Before a writer dumps to the disk, it needs to get an EXCLUSIVE lock. From RESERVED state when the writer gets PENDING(2), it affects the database in two ways:
1. It will hold onto the PENDING(2) lock no matter what
2. In PENDING(2) lock, no other connection can go from UNLOCKED to SHARED. This creates a deadlock when a previous writer has committed a transaction and it tries to get a SHARED lock. The PENDING(2) lock prevents that previous writer from entering PENDING(1) hence a deadlock.
Because of the second effect noted above, all the previous trial solutions failed. The application would retry infinitely but not get an EXCLUSIVE lock. The book suggested the use of the BEGIN IMMEDIATE command instead of the BEGIN command to start a write transaction. This prevents the deadlock situation noted in 2 above. Here is an extract from The Definitive Guide to SQLite to give more detail:
“Since you know you want to write to the database, then you need to start by issuing begin IMMEDIATE. If you get a SQLITE_BUSY, then at least you know what state you’re in. You know you can safely keep trying without holding up another connection. And once you finally do succeed, you know what state you are in then as well—RESERVED. Now you can use brute force if you have to because you are the one in the right. If you start with a begin exclusive, on the other hand, then you are assured that you won’t have any busy conditions to deal with at all. Just remember that in this case you are doing your work in EXCLUSIVE, which is not as good for concurrency as doing the work in RESERVED"
As such, the code for setAutoCommit in the Conn class was changed to use “begin immediate“. The application then needed to keep retrying if it received an
SQLITE_BUSY errors. A busy handler was created that returns 1 (i.e. try again). This wasn’t enough since the application still received the
SQLITE_BUSY exceptions even when trying to set auto commit to false. Seems it has to do with SQLite internal detecting a deadlock and just throwing an exception. Changes needed to be made to the org.sqlite.PrepStmt and org.sqlite.Conn classes. The methods:
commit, were wrapped in
try-catch blocks to keep retrying if either an
SQLITE_BUSY exception were thrown. The source code for this is in
With that solution implemented, the server now knows how to handle a locked database file. Using begin immediate has some side effects though and as such some unit tests in the JDBC driver fail.
The source code for the SQLite adapter is shared on bitbucket.org/cwdesautels/nexj-express-sqlite-adapter in the ‘sqlite’ branch. Check out the
SQLite* classes in the package
The source code to build a customized Xerial SQLite JDBC driver and SQLite shell can be found at bitbucket.org/gbatumbya/nexj-express-sqlite. (This repo use subrepos to track changes to gbatumbya/sqlite and gbatumbya/sqlite-jdbc)