use of org.apache.hadoop.hive.metastore.api.TxnOpenException in project hive by apache.
the class TestTxnHandler method testUnlockWithTxn.
@Test
public void testUnlockWithTxn() throws Exception {
LOG.debug("Starting testUnlockWithTxn");
// Test that attempting to unlock locks associated with a transaction
// generates an error
long txnid = openTxn();
LockComponent comp = new LockComponent(LockType.SHARED_WRITE, LockLevel.DB, "mydb");
comp.setTablename("mytable");
comp.setPartitionname("mypartition=myvalue");
comp.setOperationType(DataOperationType.DELETE);
List<LockComponent> components = new ArrayList<LockComponent>(1);
components.add(comp);
LockRequest req = new LockRequest(components, "me", "localhost");
req.setTxnid(txnid);
LockResponse res = txnHandler.lock(req);
long lockid = res.getLockid();
try {
txnHandler.unlock(new UnlockRequest(lockid));
fail("Allowed to unlock lock associated with transaction.");
} catch (TxnOpenException e) {
}
}
use of org.apache.hadoop.hive.metastore.api.TxnOpenException in project hive by apache.
the class TxnHandler method unlock.
/**
* This would have been made simpler if all locks were associated with a txn. Then only txn needs to
* be heartbeated, committed, etc. no need for client to track individual locks.
* When removing locks not associated with txn this potentially conflicts with
* heartbeat/performTimeout which are update/delete of HIVE_LOCKS thus will be locked as needed by db.
* since this only removes from HIVE_LOCKS at worst some lock acquire is delayed
*/
@RetrySemantics.Idempotent
public void unlock(UnlockRequest rqst) throws TxnOpenException, MetaException {
try {
Connection dbConn = null;
Statement stmt = null;
long extLockId = rqst.getLockid();
try {
/**
* This method is logically like commit for read-only auto commit queries.
* READ_COMMITTED since this only has 1 delete statement and no new entries with the
* same hl_lock_ext_id can be added, i.e. all rows with a given hl_lock_ext_id are
* created in a single atomic operation.
* Theoretically, this competes with {@link #lock(org.apache.hadoop.hive.metastore.api.LockRequest)}
* but hl_lock_ext_id is not known until that method returns.
* Also competes with {@link #checkLock(org.apache.hadoop.hive.metastore.api.CheckLockRequest)}
* but using SERIALIZABLE doesn't materially change the interaction.
* If "delete" stmt misses, additional logic is best effort to produce meaningful error msg.
*/
dbConn = getDbConn(Connection.TRANSACTION_READ_COMMITTED);
stmt = dbConn.createStatement();
// hl_txnid <> 0 means it's associated with a transaction
String s = "DELETE FROM \"HIVE_LOCKS\" WHERE \"HL_LOCK_EXT_ID\" = " + extLockId + " AND (\"HL_TXNID\" = 0 OR" + " (\"HL_TXNID\" <> 0 AND \"HL_LOCK_STATE\" = '" + LOCK_WAITING + "'))";
// (hl_txnid <> 0 AND hl_lock_state = '" + LOCK_WAITING + "') is for multi-statement txns where
// some query attempted to lock (thus LOCK_WAITING state) but is giving up due to timeout for example
LOG.debug("Going to execute update <" + s + ">");
int rc = stmt.executeUpdate(s);
if (rc < 1) {
LOG.info("Failure to unlock any locks with extLockId={}.", extLockId);
dbConn.rollback();
Optional<LockInfo> optLockInfo = getLockFromLockId(dbConn, extLockId);
if (!optLockInfo.isPresent()) {
// didn't find any lock with extLockId but at ReadCommitted there is a possibility that
// it existed when above delete ran but it didn't have the expected state.
LOG.info("No lock in " + LOCK_WAITING + " mode found for unlock(" + JavaUtils.lockIdToString(rqst.getLockid()) + ")");
// bail here to make the operation idempotent
return;
}
LockInfo lockInfo = optLockInfo.get();
if (isValidTxn(lockInfo.txnId)) {
String msg = "Unlocking locks associated with transaction not permitted. " + lockInfo;
// if a lock is associated with a txn we can only "unlock" if if it's in WAITING state
// which really means that the caller wants to give up waiting for the lock
LOG.error(msg);
throw new TxnOpenException(msg);
} else {
// we didn't see this lock when running DELETE stmt above but now it showed up
// so should "should never happen" happened...
String msg = "Found lock in unexpected state " + lockInfo;
LOG.error(msg);
throw new MetaException(msg);
}
}
LOG.debug("Successfully unlocked at least 1 lock with extLockId={}", extLockId);
dbConn.commit();
} catch (SQLException e) {
LOG.error("Unlock failed for request={}. Exception msg: {}", rqst, getMessage(e));
rollbackDBConn(dbConn);
checkRetryable(e, "unlock(" + rqst + ")");
throw new MetaException("Unable to update transaction database " + JavaUtils.lockIdToString(extLockId) + " " + StringUtils.stringifyException(e));
} finally {
closeStmt(stmt);
closeDbConn(dbConn);
}
} catch (RetryException e) {
unlock(rqst);
}
}
use of org.apache.hadoop.hive.metastore.api.TxnOpenException in project hive by apache.
the class Cleaner method removeFiles.
private boolean removeFiles(String location, long minOpenTxnGLB, CompactionInfo ci, boolean dropPartition) throws MetaException, IOException, NoSuchObjectException, NoSuchTxnException {
if (dropPartition) {
LockRequest lockRequest = createLockRequest(ci, 0, LockType.EXCL_WRITE, DataOperationType.DELETE);
LockResponse res = null;
try {
res = txnHandler.lock(lockRequest);
if (res.getState() == LockState.ACQUIRED) {
// check if partition wasn't recreated
if (resolvePartition(ci) == null) {
return removeFiles(location, ci);
}
}
} catch (NoSuchTxnException | TxnAbortedException e) {
LOG.error(e.getMessage());
} finally {
if (res != null && res.getState() != LockState.NOT_ACQUIRED) {
try {
txnHandler.unlock(new UnlockRequest(res.getLockid()));
} catch (NoSuchLockException | TxnOpenException e) {
LOG.error(e.getMessage());
}
}
}
}
ValidTxnList validTxnList = TxnUtils.createValidTxnListForCleaner(txnHandler.getOpenTxns(), minOpenTxnGLB);
// save it so that getAcidState() sees it
conf.set(ValidTxnList.VALID_TXNS_KEY, validTxnList.writeToString());
/**
* {@code validTxnList} is capped by minOpenTxnGLB so if
* {@link AcidUtils#getAcidState(Path, Configuration, ValidWriteIdList)} sees a base/delta
* produced by a compactor, that means every reader that could be active right now see it
* as well. That means if this base/delta shadows some earlier base/delta, the it will be
* used in favor of any files that it shadows. Thus the shadowed files are safe to delete.
*
* The metadata about aborted writeIds (and consequently aborted txn IDs) cannot be deleted
* above COMPACTION_QUEUE.CQ_HIGHEST_WRITE_ID.
* See {@link TxnStore#markCleaned(CompactionInfo)} for details.
* For example given partition P1, txnid:150 starts and sees txnid:149 as open.
* Say compactor runs in txnid:160, but 149 is still open and P1 has the largest resolved
* writeId:17. Compactor will produce base_17_c160.
* Suppose txnid:149 writes delta_18_18
* to P1 and aborts. Compactor can only remove TXN_COMPONENTS entries
* up to (inclusive) writeId:17 since delta_18_18 may be on disk (and perhaps corrupted) but
* not visible based on 'validTxnList' capped at minOpenTxn so it will not not be cleaned by
* {@link #removeFiles(String, ValidWriteIdList, CompactionInfo)} and so we must keep the
* metadata that says that 18 is aborted.
* In a slightly different case, whatever txn created delta_18 (and all other txn) may have
* committed by the time cleaner runs and so cleaner will indeed see delta_18_18 and remove
* it (since it has nothing but aborted data). But we can't tell which actually happened
* in markCleaned() so make sure it doesn't delete meta above CG_CQ_HIGHEST_WRITE_ID.
*
* We could perhaps make cleaning of aborted and obsolete and remove all aborted files up
* to the current Min Open Write Id, this way aborted TXN_COMPONENTS meta can be removed
* as well up to that point which may be higher than CQ_HIGHEST_WRITE_ID. This could be
* useful if there is all of a sudden a flood of aborted txns. (For another day).
*/
// Creating 'reader' list since we are interested in the set of 'obsolete' files
ValidReaderWriteIdList validWriteIdList = getValidCleanerWriteIdList(ci, validTxnList);
LOG.debug("Cleaning based on writeIdList: {}", validWriteIdList);
return removeFiles(location, validWriteIdList, ci);
}
Aggregations