What are Deadlocks?
Here is an example:
Transaction 1 locks some rows in table 1, and transaction 2 locks some rows in table 2. The next step of transaction 1 is to lock rows in table 2 and transaction 2 needs to lock rows in table 1.
We just described a deadlock situation. Without deadlock handling, both transactions would wait forever, or until they are aborted. On Teradata, deadlocks can happen on one AMP (local deadlock) or across different AMPs (global deadlock), and for various reasons.
Luckily, Teradata uses a queuing strategy which is serializing locking requests to avoid deadlocks.
There was a change in the naming convention for this locking strategy in Teradata 15.10.
Until Teradata 14.10, this locking strategy was called “pseudo table locks” independently if the lock was on row hash level (for dictionary tables) or table level.
Since Teradata 15.10, table level and partition locking (a new feature) are called “proxy locking”, and rows hash locking on dictionary tables is called “pseudo table locking.”
In my opinion, the strategy is still the same (but what's new since Teradata 15.10 is partition locking). Just the wording changed.
The “Proxy” or “Pseudo Table” Lock Strategy
Without a proper locking strategy, and two transactions asking for a write lock on the same table, it could happen that the first transaction gets the lock for the table on some of the AMPs, and the second query takes the locks on another set of AMPs.
None of the transactions would be able to finish its task. Both requests would wait forever (for completeness: there is an NOWAIT lock modifier available, which aborts the request if the lock can't be obtained immediately).
A typical global deadlock situation (involving several AMPs).
The “proxy lock” or “pseudo table” strategy avoids such deadlocks, by serializing the requests.
For sure many of you all have seen the term “to prevent global deadlock” when explaining a query (the explain plan is from a Teradata 15.10 system):
Explain UPDATE CUSTOMER SET AGE = 40;
Teradata Deadlock Prevention in Action
Here is an example, which shows the deadlock handling strategy in action:
Two update statements execute at almost the same time (‘Update 1' and ‘Update 2'), and they want to update the same table. Each update intends to have the write lock:
UPDATE Customer SET Age=40 ;
UPDATE Customer SET Gender = ‘M';
“Pseudo table” or “Proxy” locking ensures that each request has to get the pseudo lock on a reserved rowhash, before obtaining the required lock.
For each table, “proxy” or “pseudo table” locking defines an AMP which is the gatekeeper to the locks. The gatekeeper AMP for each table is found by hashing its “table id” value.
Hashing happens in the same way like primary index hashing. By hashing the “table id”, the gatekeeper AMP is found.
As the hashing algorithm is stable, the rows hash for a specific “table id” always is the same, and as long as the system configuration is not changed, there is a 1:1 relation between table and gatekeeper AMP.
In our example, there are two update statements which want a write lock on the same table.
Assuming that “update 1” is slightly faster than “update 2”, it will get the “proxy” or “pseudo table” lock on AMP 2, which is the gatekeeper for the table “Customer.” “Update 2” has to wait in the “gatekeeper” queue of AMP 2 and will be next, as soon as “update 2” finished and released the write lock.
Not all kind of deadlocks can be avoided with the above-described strategy. It only works if both participating transactions work with table locks. If one or both request use row-hash locking, deadlocks still can happen.
Furthermore, deadlock detection takes time. Teradata checks by default for global deadlocks every four minutes. Local deadlocks are searched every 30 seconds.
Usually, these times are ok, but in some particular cases, you might want to decrease global deadlock detection intervals.
One of my clients uses a lot of join indexes, which are causing many global deadlocks (the join indexes are needed for primary index access, i.e. row hash locks are used).
The join indexes are utilized by tactical workload requests, to keep execution times below a couple of seconds.
Having global deadlock detection set to four minutes is counterproductive in this case.