Major Description
From Adar:
This code:
// Test sending edits too fast
session.setFlushMode(KuduSession.FlushMode.AUTO_FLUSH_BACKGROUND);
session.setMutationBufferSpace(10);
// The buffer has a capacity of 10, we insert 21 rows, meaning we fill the first one,
// force flush, fill a second one before the first one could come back,
// and the 21st row will be sent back.
boolean gotException = false;
for (int i = 50; i < 71; i++) {
try {
session.apply(createInsert(i));
} catch (PleaseThrottleException ex) {
gotException = true;
assertEquals(70, i);
// Wait for the buffer to clear
ex.getDeferred().join(DEFAULT_SLEEP);
session.apply(ex.getFailedRpc());
session.flush().join(DEFAULT_SLEEP);
}
}
assertTrue(gotException); <--
assertEquals(21, countInRange(50, 71));
If the client is running particularly slowly, isn't it possible for the background flush to finish before the 11th (or possibly 21st) row is inserted? Then we won't throw a PleaseThrottleException.
It's indeed possible and it happened once. We need a way to block the background flushing from happening, maybe through mocking.