Friday, February 27, 2015

Salt: Combining ACID and BASE in a Distributed Database

This paper appeared in OSDI'14. The authors are Chao Xie, Chunzhi Su, Manos Kapritsos, Yang Wang, Navid Yaghmazadeh, Lorenzo Alvisi, and Prince Mahajan, all from The University of Texas at Austin. Here you can watch a video of the OSDI presentation of the paper, and also find the presentation slides and the paper available as open access.  USENIX knows how to do conferences right.

Dropping ACID

ACID (Atomicity, Consistency, Isolation, Durability) approach provides ease-of-programming through its simple transaction abstraction, but loses on the performance. BASE (Basically-Available, Soft state, Eventually consistent) approach, popularized by the NoSQL systems, provides good performance (low-latency, high-throughput, and scalability), but loses on the ease-of-programming due to increased complexity of concurrent execution in distributed systems.

This paper introduces Salt, which aims to find a best-of-both-worlds middle ground: to provide high performance like BASE with modest programming effort like ACID.

Less is more

Salt approach is motivated by the following observation (which is an instance of the Pareto principle). I am quoting from the paper:
"When an application outgrows the performance of an ACID implementation, it is often because of the needs of only a handful of transactions: most transactions never test the limits of what ACID can offer. Numerous applications [2, 4, 5, 10, 11] demonstrate this familiar lopsided pattern: few transactions are performance-critical, while many others are either lightweight or infrequent." 
Of course, a better quantification than saying "numerous" or "many" applications would make the argument for the Salt approach stronger. However, such an extensive study may take several months, and is unfair to ask for a conference paper submission.

It is tempting to increase the concurrency of those critical ACID transactions by splitting them into smaller mini ACID transactions. However, doing so may not be safe as you would lose Atomicity (which frees you from worrying about intermediate states during failures) and Isolation (which regulates which states can be accessed when transactions execute concurrently). The paper calls this a stark choice and cites the YouTube clip from Butch Cassidy to further drive this point home (I am not kidding).

The paper's proposal is this. Instead of worrying about all possible Atomicity & Isolation complications with all other ACID transactions, Salt categorizes transactions into ACID transactions and a small number of BASE transactions, and lets you worry only about checking Atomicity & Isolation complications between the BASE transactions.

Of course there is an unmentioned complication here. You also have to ensure that these BASE transactions leave the database in consistent state. This is not an easy task as well, and opens up Salt's ease-of-programming to debate. On the other hand, you would have to worry about this for all operations if you implement a full-BASE solution.

The Salt approach

Actually Salt's BASE transactions approach gets its inspiration from the ACID nested transactions. The paper says:
"Nested transactions is an abstraction originally introduced to offer, for long-running transactions, atomicity at a finer granularity than isolation. Our purpose in introducing BASE transactions is similar in spirit to that of traditional nested transactions: both abstractions aim at gently loosening the coupling between atomicity and isolation. The issue that BASE transactions address, however, is the flip side of the one tackled by nested transactions: this time, the challenge is to provide isolation at a finer granularity, without either drastically escalating the complexity of reasoning about the application, or shattering atomicity."

By providing isolation at a finer granularity, Salt allows BASE transactions to achieve high concurrency by observing each other's internal states, without affecting the isolation guarantees of ACID transactions. A BASE transaction is made up of multiple mini-transactions, called alkaline subtransactions. (Yes, the paper makes several Chemistry puns.) The committed state of an alkaline subtransaction is observable by other BASE or alkaline subtransactions. On the other hand, the committed state of an alkaline subtransaction is not observable by other ACID transactions until the parent BASE transaction commits.  As such, Salt guarantees that, when BASE and ACID transactions execute concurrently, ACID transactions retain, with respect to all other transactions (whether BASE, alkaline, or ACID), the same isolation guarantees they used to enjoy in a purely ACID environment.

The proverbial Bank example

When you talk about ACID versus BASE the classical example is the bank money transfer example. The paper gives a pure ACID, a pure BASE, and a Salt implementation of this example.

I wonder, which implementation is closer to the way the money transfers are actually handled in real banks. If someone familiar with how banks handle the money transfer can weigh in on this, that may shed a light on this debate.  My money is on the pure BASE approach.

Something about this bank example is odd though. The total-balance operation is an unrealistic big/monolithic operation. Why did that have to be a transaction? The total-balance operation could be done with a consistent/timed snapshot, no? Then why did it have to be a transactions spanning/freezing the entire system?

Now let's check the three implementations more closely. The implementation in Figure 1.a (ACID) uses only one type of locks, ACID locks. Figure 1.b (BASE) uses only one type of locks, local locks. BASE programming looks like classical distributed system programming, it tries to do everything with local asynchronous actions as much as possible. Figure 2 (Salt) uses Alkaline, Saline, and ACID locks.

Why does Salt need 3 different locks? Because the first alkaline transaction is special, as it determines whether the entire BASE transaction it is included in will succeed or not. If that first alkaline subtransaction succeeds, then comes the saline-lock to make the intermediate states in the rest of the BASE transaction available to other BASE and alkaline transactions. But then, this is something I am still not clear on. Salt attributes significance to the first alkaline subtransaction, so its design should also be special. But I am unclear about what should go in that subtransaction? Are there guidelines to design that first subtransaction so it captures the successful completion of the rest of the BASE transaction?


Like any OSDI/SOSP paper worth its salt (see, I can also make "Salt" puns :-), the paper includes a strong evaluation section. The authors implemented a Salt prototype by modifying MySQL Cluster. They evaluated the prototype using the TPC-C benchmark, which consists of 5 types of transactions: new-order (43.5%), payment (43.5%), stock-level (4.35%), order-status (4.35%), and delivery (4.35%). The results show that by BASE-ifying 2 of these transactions, Salt achieves 80% of maximum throughput of a full BASE implementation.

Notice the left side of Figure 7, which shows that ACID version faster than the Salt version under low load. The paper doesn't discuss this but, if we had both ACID and BASE version of transactions, it may be possible to hot-swap between these two versions to optimize the throughput of the system depending on the current system load.

Related work

The paper fails to cite several papers that proposed orthogonal approaches to deal with the same problem: improving the performance of ACID transactions, while keeping the ease-of-programmability.

The Red-Blue Actions paper (OSDI'12) considered the problem of reducing the granularity of ACID transactions by using a generator operation and shadow operation and categorizing the operations  as red (consistency critical) and blue (fast and eventually consistent). That paper also uses a Bank example.

Then there are work from the Berkeley group on this problem. I had summarized the Invariant-based coordination avoidance paper earlier. There is also the Highly Available Transactions paper  (VLDB'13) where they  look at ACID isolation level relaxing for improving availability.


The paper proposes the Salt approach assuming the ACID implementation of the application is already present. The application for this is if you already have an ACID system, you can Salt it to fine-tune the performance. Another approach would be to start with a BASE implementation and to increase atomicity/isolation for a couple operations. Maybe for some applications, it will be easier for going from BASE to Salt: Starting with BASE and then adding Alkaline subtransactions and ACID transactions (if necessary).

The Salt approach do require a learning curve and can be tricky. But how do we argue the complexity of the programming effort needed for one approach (say Salt) versus the other (say pure BASE) in an objective manner? One can objectively show performance improvements with evaluations, but it is harder to argue for "ease of programming" in an objective/quantitative manner.

There can also be some complications with the early-committing (after the first alkaline subtransaction) in BASE transactions. This question came up in the OSDI presentation. If deadlock occurs the MySQL Cluster approach is to use timouts and roll the transactions back. In this case, Salt throws exceptions which the developer need to address to re-achive consistency and atomicity.

No comments: