I am encountering very infrequent yet annoying SQL deadlocks on a .NET 2.0 webapp running on top of MS SQL Server 2005. In the past, we have been dealing with the SQL deadlocks in the very empirical way - basically tweaking the queries until it work.
Yet, I found this approach very unsatisfactory: time consuming and unreliable. I would highly prefer to follow deterministic query patterns that would ensure by design that no SQL deadlock will be encountered - ever.
For example, in C# multithreaded programming, a simple design rule such as the locks must be taken following their lexicographical order ensures that no deadlock will ever happen.
Are there any SQL coding patterns guaranteed to be deadlock-proof?
Writing deadlock-proof code is really hard. Even when you access the tables in the same order you may still get deadlocks [1]. I wrote a post on my blog that elaborates through some approaches that will help you avoid and resolve deadlock situations.
If you want to ensure two statements/transactions will never deadlock you may be able to achieve it by observing which locks each statement consumes using the sp_lock system stored procedure. To do this you have to either be very fast or use an open transaction with a holdlock hint.
Notes:
Any SELECT statement that needs more than one lock at once can deadlock against an intelligently designed transaction which grabs the locks in reverse order.
Zero deadlocks is basically an incredibly costly problem in the general case because you must know all the tables/obj that you're going to read and modify for every running transaction (this includes SELECTs). The general philosophy is called ordered strict two-phase locking (not to be confused with two-phase commit) (http://en.wikipedia.org/wiki/Two_phase_locking ; even 2PL does not guarantee no deadlocks)
Very few DBMS actually implement strict 2PL because of the massive performance hit such a thing causes (there are no free lunches) while all your transactions wait around for even simple SELECT statements to be executed.
Anyway, if this is something you're really interested in, take a look at SET ISOLATION LEVEL in SQL Server. You can tweak that as necessary. http://en.wikipedia.org/wiki/Isolation_level
For more info, see wikipedia on Serializability: http://en.wikipedia.org/wiki/Serializability
That said -- a great analogy is like source code revisions: check in early and often. Keep your transactions small (in # of SQL statements, # of rows modified) and quick (wall clock time helps avoid collisions with others). It may be nice and tidy to do a LOT of things in a single transaction -- and in general I agree with that philosophy -- but if you're experiencing a lot of deadlocks, you may break the trans up into smaller ones and then check their status in the application as you move along. TRAN 1 - OK Y/N? If Y, send TRAN 2 - OK Y/N? etc. etc
As an aside, in my many years of being a DBA and also a developer (of multiuser DB apps measuring thousands of concurrent users) I have never found deadlocks to be such a massive problem that I needed special cognizance of it (or to change isolation levels willy-nilly, etc).
There is no magic general purpose solution to this problem that work in practice. You can push concurrency to the application but this can be very complex especially if you need to coordinate with other programs running in separate memory spaces.
General answers to reduce deadlock opportunities:
Basic query optimization (proper index use) hotspot avoidanant design, hold transactions for shortest possible times...etc.
When possible set reasonable query timeouts so that if a deadlock should occur it is self-clearing after the timeout period expires.
Deadlocks in MSSQL are often due to its default read concurrency model so its very important not to depend on it - assume Oracle style MVCC in all designs. Use snapshot isolation or if possible the READ UNCOMMITED isolation level.
I believe the following useful read/write pattern is dead lock proof given some constraints:
Constraints:
One table
An index or PK is used for read/write so engine does not resort to table locks.
A batch of records can be read using a single SQL where clause.
Using SQL Server terminology.
Write Cycle:
All writes within a single "Read Committed" transaction.
The first update in the transaction is to a specific, always-present record
within each update group.
Multiple records may then be written in any order. (They are "protected"
by the write to the first record).
Read Cycle:
The default read committed transaction level
No transaction
Read records as a single select statement.
Benefits:
Secondary write cycles are blocked at the write of first record until the first write transaction completes entirely.
Reads are blocked/queued/executed atomically between the write commits.
Achieve transaction level consistency w/o resorting to "Serializable".
I need this to work too so please comment/correct!!
As you said, always access tables in the same order is a very good way to avoid deadlocks. Furthermore, shorten your transactions as much as possible.
Another cool trick is to combine 2 sql statements in one whenever you can. Single statements are always transactional. For example use "UPDATE ... SELECT" or "INSERT ... SELECT", use "##ERROR" and "##ROWCOUNT" instead of "SELECT COUNT" or "IF (EXISTS ...)"
Lastly, make sure that your calling code can handle deadlocks by reposting the query a configurable amount of times. Sometimes it just happens, it's normal behaviour and your application must be able to deal with it.
In addition to consistent sequence of lock acquisition - another path is explicit use of locking and isolation hints to reduce time/resources wasted unintentionally acquiring locks such as shared-intent during read.
Something that none has mentioned (surprisingly), is that where SQL server is concerned many locking problems can be eliminated with the right set of covering indexes for a DB's query workload. Why? Because it can greatly reduce the number of bookmark lookups into a table's clustered index (assuming it's not a heap), thus reducing contention and locking.
If you have enough design control over your app, restrict your updates / inserts to specific stored procedures and remove update / insert privileges from the database roles used by the app (only explicitly allow updates through those stored procedures).
Isolate your database connections to a specific class in your app (every connection must come from this class) and specify that "query only" connections set the isolation level to "dirty read" ... the equivalent to a (nolock) on every join.
That way you isolate the activities that can cause locks (to specific stored procedures) and take "simple reads" out of the "locking loop".
Quick answer is no, there is no guaranteed technique.
I don't see how you can make any application deadlock proof in general as a design principle if it has any non-trivial throughput. If you pre-emptively lock all the resources you could potentially need in a process in the same order even if you don't end up needing them, you risk the more costly issue where the second process is waiting to acquire the first lock it needs, and your availability is impacted. And as the number of resources in your system grows, even trivial processes have to lock them all in the same order to prevent deadlocks.
The best way to solve SQL deadlock problems, like most performance and availability problems is to look at the workload in the profiler and understand the behavior.
Not a direct answer to your question, but food for thought:
http://en.wikipedia.org/wiki/Dining_philosophers_problem
The "Dining philosophers problem" is an old thought experiment for examining the deadlock problem. Reading about it might help you find a solution to your particular circumstance.
Related
I have set of validations which decides record to be inserted into the database with valid status code, the issue we are facing is that many users are making requests at the same time and middle of one transaction another transaction comes and both are getting inserted with valid status, which it shouldn't. it should return an error that record already exists which can be easily handled by a simple query but at specific scenarios we are allowing them to insert duplicates, I have tried sp_getapplock which is solving my problem but it is compromising performance big time. Are there any optimal ways to handle concurrent requests?
Thanks.
sp_getapplock is pretty much the befiest and most arbitrary lock you can take. It functions more like the lock keyword does in OOO programming. Basically you name a resource, give it a scope (proc or transaction), then lock it. Pretty much nothing can bypass that lock, which is why it's solved your race conditions. It's also probably mad overkill for what you're trying to do.
The first code/architecture idea that comes to mind is to restructure this table. I'm going to assume you have high update volumes or you wouldn't be running into these violations. You could simply use a try/catch block, and have the catch block retry on a PK violation. Clumsy, but might just do the trick.
Next, you could consider altering the structure of the table which receives this stream of updates throughout the day. Make this table primary keyed off an identity column, and pretty much nothing else. Inserts will be lightning fast, so any blockage will be negligible. You can then move this data in batches into a table better suited for batch processing (as opposed to trying to batch-process in real time)
There are also a whole range of transaction isolation settings which adjust SQL's regular locking system to support different variants (whether at the batch level, or inline via query hints. I'd read up on those, but you might consider looking at Serialized isolation. Various settings will enforce different runtime rules to fit your needs.
Also be sure to check your transactions. You probably want to be locking the hell out of this table (and potentially during some other action) but once that need is gone, so should the lock.
Let's say I have a table with 1,000,000 rows and running a SELECT * FROM TableName on this table takes around 10 seconds to return the data.
Without a NOLOCK statement (putting to one side issues around dirty reads) would this query lock the table for 10 seconds meaning that no other process could read or write to the table?
I am often informed by DBAs then when querying live data to diagnose data issues I should use NOLOCK to ensure that I don't lock the table which may cause issues for users. Is this true?
The NOLOCK table hint will cause that no shared locks will be taken for the table in question; same with READUNCOMMITTED isolation level, but this time applies not to a single table but rather to everything involved. So, the answer is 'no, it won't lock the table'. Note that possible schema locks will still be held even with readuncommitted.
Not asking for shared locks the read operation will potentially read dirty data (updated but not yet committed) and non-existent data (updated, but rolled back), transactionally inconsistent in any case, and migh even skip the whole pages as well (in case of page splits happening simultaneously with the read operation).
Specifying either NOLOCK or READUNCOMMITTED is not considered a good practice. It will be faster, of course. Just make sure you're aware of consequences.
Also, support for these hints in UPDATE and DELETE statements will be removed in a future version, according to docs.
Even with NOLOCK, SQL Server can choose to override and obtain a lock, nevertheless. It is called a QUERY **HINT** for a reason. The sure fire way of avoiding locks would be to use SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED in the beginning of the session.
Your query will try to obtain a table level lock as you're asking for all the data. So yes it will be held for the duration of the query and all other processes trying to acquire Exclusive locks on that table with be put on wait queues. Processes that are also trying to read from that table will not be blocked.
Any diagnosis performed on a live system should done with care and using the NOLOCK hint will allow you to view data without creating any contention for users.
EDIT: As pointed out update locks are compatible with shared locks. So won't be blocked by the read process.
Is it possible in relational databases for these two statements to deadlock? I'm trying to simplify my question and example -- please just assume that these selects, which I think would normally only require sharable read-locking, now require exclusive read locks:
Concurrent Connection 1:
SELECT {...}
FROM A
JOIN B ON {...}
Concurrent Connection 2:
SELECT {...}
FROM B
JOIN A ON {...}
That is, does the ordering of the joins matter? Are single statements in SQL atomic? Is A locked first and then B in the first statement and B locked first and then A in the second statement?
I think not - My gut tells me that two single statements like this cannot deadlock, no matter how complex. I believe that a statement is analyzed as a whole and that the resources requiring locking are locked using some deterministic global order (i.e. alphabetically). But I need more than a gut feeling on this - I can't think of a way to prove it and I can't find it documented.
I'm interested in MS SQL 2005, but I don't think the question is implementation specific.
Secondarily: As it relates to MS SQL, I'd also want to know that Common Table Expressions also have this guarantee - that CTEs are mostly a syntactic benefit (+recursion), consolidated into a traditional single statement by the engine.
SELECTs cannot deadlock with other SELECT, because they only acquire shared locks. You say that we should consider that these SELECTs now 'require exclusive read locks', but this is not possible for us to consider because 1) there is no such thing as an exlusive read lock and 2) reads don't acquire exclusive locks.
But you do pose a more general question, whether simple statements can deadlock. The answer is a definite, resounding YES. Locks are acquired at execution, not analyzed upfront and sorted then acquired in some order. It would be impossible for the engine to know upfront the needed locks because they depend on the actual data in on-disk, and to read the data the engine needs to ... lock the data.
Deadlocks between simple statements (SELECt vs. UPDATE or SELECT vs. DELETE) due to different index access order are quite common and very easy to investigate, diagnose and fix. But note that there is always a write operation involved, as reads cannot block each other. For this discussion, adding a UPDLOCK or XLOCK hint to a SELECT should be considered a write. You don't even need a JOIN, a secondary index may well introduce the access order problem leading to deadlock, see Read/Write Deadlock.
And finally, writing SELECT FROM A JOIN B or writing SELECT FROM B JOIN A is completely irrelevant. The query optimizer is free to rearrange the access order as it sees fit, the actual text of the query does not impose the order of execution in any way.
Updated
How then can we construct a general
strategy toward a READ COMMITTED
"multiple entity" database that
doesn't deadlock?
I'm afraid there is no cookie-cutter recipe. The solution will depend from case to case. Ultimately, in database applications deadlocks are a fact of life. I understand this may sound absurd, as in 'we landed on the Moon but we can't write a correct database application', but there are strong factors at play which pretty much guarantee that applications will eventually encounter deadlocks. Lucky deadlocks are the easiest to deal with errors, simple read again the state, apply the logic, re-write the new state. Now that being said, there are some good practices that can dramatically reduce the frequency of deadlocks, down to the point they are all but vanished:
Try to have a consistent access pattern for Writes. Have clearly defined rules stating things such as 'a transaction shall always tables in this order: Customers -> OrderHeaders -> OrderLines.' Note that the order has to be obeyed inside a transaction. Basically, rank all tables in your schema and specify that all updates must occur in ranking order. This eventually boils down to code discipline of the individual contributor writing the code, as it has to ensure it writes is update sin the proper order inside a transaction.
Reduce the duration of writes. The usual wisdom goes as this: at the beginning of the transaction do all the reads (read the existing state), then process the logic and compute new values, then write all updates at the end of transaction. Avoid a pattern like 'read->write->logic->read->write', instead do 'read->read->logic->write->write'. Of course, the true craftsmanship consist in how to deal with actual, real, individual cases when apparently one must have to do writes mid-transaction. A special note here must be said about a specific type of transaction: those driven by a queue, which by very definition start their activity by dequeueing (= a write) from the queue. These applications were always notoriously difficult to write and prone to errors (specially deadlocks), luckily there are ways to do it, see Using tables as Queues.
Reduce the amount of reads. Table scans are the most prevalent cause of deadlocks. Proper indexing will not only eliminate the deadlocks, but may also boost performance in the process.
Snapshot isolation. This is the closest thing you'll get to a free lunch in regard to avoiding deadlocks. I intentionally put it last, because it may mask other problems (like improper indexing) instead of fixing them.
Trying to solve this problem with a LockCustomerByXXX approach I'm afraid doesn't work. Pessimistic locking doesn't scale. Optimistic concurrency updates are the way to go if you want to have any sort of decent performance.
As far as I know, you are correct: the SQL engine figures out what it will need to do (probably as it parses the query), locks all required resources, executes the query, and then unlocks them.
Reads won't deadlock each other. You must have some write going on as well.
You can do things to reduce the number of deadlocks. For example, insert only at the end of a clustered index on platforms that support row locking and avoid updating records. Ah hah, now Facebook's UI makes more sense.
It's sometimes easier to handle the deadlocks than avoid them. The server will fail and report back, and you can retry.
I'm using MySql 5.x and in my environment, I have a table with the name CALLS.
Table CALLS has a column status which takes an enum {inprogress, completed}.
I want reads/updates of the table to be row-locked, so:
SET TRANSACTION ISOLATION LEVEL SERIALIZABLE;
SET AUTOCOMMIT = 0;
SELECT amount from CALLS where callId=1213 FOR UPDATE;
COMMIT
Basically I'm doing a FOR UPDATE even in situations whereby I only need to read the amount and return. I find that this allow me to ensure that reads/updates are prevented from interfering from each other. However I've been told this will reduce the concurrency of the app.
Is there anyway to achieve the same transaction consistency without incurring locking overheads ? Thanks.
Disclaimer: MySQL is generally full of surprises, so the following could be untrue.
What you are doing doesn't make any sense to me: You are committing after the SELECT, which should break the lock. So in my opinion, your code shouldn't really incur any significant overhead; but it doesn't give you any consistency improvements, either.
In general, SELECT FOR UPDATE can be a very sound and reasonable way to ensure consistency without taking more locks than are really needed. But of course, it should only be used when needed. Maybe you should have different code paths: One (using FOR UPDATE) used when the retrieved value is used in a subsequent change-operation. And another one (not using FOR UPDATE) used when the value doesn't have to be protected from changes.
What you've implemented there--in case you weren't familiar with it--is called pessimistic locking. You're sacrificing performance for consistency, which is sometimes a valid choice. In my professional experience, I've found pessimistic locking to be far more of a hindrance than a help.
For one thing, it can lead to deadlock.
The (better imho) alternative is optimistic locking, where you make the assumption that collisions occur infrequently and you simply deal with them when they happen. You're doing your owrk in a transaction so a collision shouldn't leave your data in an inconsistent state.
Here's more information on optimistic locking in a Java sense but the ideas are applicable to anything.
A DBA that my company hired to troubleshoot deadlock issues just told me that our OLTP databases locking problems will improve if we set the transaction level to READ COMMITTED from READ UNCOMMITTED.
Isn't that just 100% false? READ COMMITTED will cause more locks, correct?
More Details:
Our data is very "siloed" and user specific. 99.9999999 % of all user interactions work with your own data and our dirty read scenarios, if they happen, can barely effect what the user is trying to do.
Thanks for all the answers, the dba in question ended up being useless, and we fixed the locking issues by adding a single index.
I regret that I didn't specify the locking problems were occurring for update statements and not regular selects. From my googing the two different query types have distinct solutions when dealing with locking issues.
That does sound like a bit of a rash decision, however without all the details of your environment it is difficult to say.
You should advise your DBA to consider the use of SQL Server's advanced isolation features, i.e. the use of Row Versioning techniques. This was introduced to SQL Server 2005 to specifically address issues with OLTP database that experience high locking.
The following white paper contains quite complicated subject matter but it is a must read for all exceptional DBA's. It includes example of how to use each of the additional isolation levels, in different types of environments i.e. OLTP, Offloaded Reporting Environment etc.
http://msdn.microsoft.com/en-us/library/ms345124.aspx
In summary it would be both foolish and rash to modify the transaction isolation for all of your T-SQL queries without first developing a solid understanding of how the excessive locking is occuring within your environment.
I hope this helps but please let me know if you require further clarification.
Cheers!
Doesn't it depend on what your problem is: for example if your problem is a deadlock, mightn't an increase in the locking level cause an earlier acquisition of locks and therefore a decreased possibility of deadly embrace?
It the data is siloed and you are still getting deadlocks then you may simply need to add rowlock hints to the queries that are causing the problem so that the locks are taken at the row level and not the page level (which is the default).
READ UNCOMMITTED will reduce the number of locks if you are locking data because of SELECT statements. If you are locking data because of INSERT, UPDATE and DELETE statements then changing the isolation level to READ UNCOMMITTED won't do anything for you. READ UNCOMMITTED has the same effect as adding WITH (NOLOCK) to your queries.
This sounds scary. Do you really want to just change these parameters to avoid deadlocks? Maybe the data needs to be locked?
That said, it could be that the DBA is referring to the new (as of SQL Server 2005) READ COMMITTED SNAPSHOT that uses row versioning and can eliminate some kinds of deadlocks.
http://www.databasejournal.com/features/mssql/article.php/3566746/Controlling-Transactions-and-Locks-Part-5-SQL-2005-Snapshots.htm