Adv DB: CAP and ACID

Transactions

A transaction is a set of operations/transformations to be carried out on a database or relational dataset from one state to another.  Once completed and validated to be a successful transaction, the ending result is saved into the database (Panda et al, 2011).  Both ACID and CAP (discussed in further detail) are known as Integrity Properties for these transactions (Mapanga & Kadebu, 2013).

 Mobile Databases

Mobile devices have become prevalent and vital for many transactions when the end-user is unable to access a wired connection.  Since the end-user is unable to find a wired connection to conduct their transaction their device will retrieve and save information on transaction either on a wireless connection or disconnected mode (Panda et al, 2011).  A problem with a mobile user accessing and creating a transaction with databases, is the bandwidth speeds in a wireless network are not constant, which if there is enough bandwidth connection to the end-user’s data is rapid, and vice versa.  There are a few transaction models that can efficiently be used for mobile database transactions: Report and Co-transactional model; Kangaroo transaction model; Two-Tiered transaction model; Multi-database transaction model; Pro-motion transaction model; and Toggle Transaction model.  This is in no means an exhaustive list of transaction models to be used for mobile databases.

According to Panda et al (2011), in a Report and Co-transactional Model, transactions are completed from the bottom-up in a nested format, such that a transaction is split up between its children and parent transaction.  The child transaction once successfully completed then feeds that information up to the chain until it reaches the parent.  However, not until the parent transaction is completed is everything committed.  Thus, a transaction can occur on the mobile device but not be fully implemented until it reaches the parent database. The Kangaroo transaction model, a mobile transaction manager collects and accepts transactions from the end-user, and forwards (hops) the transaction request to the database server.  Transaction made in this model is done by proxy in the mobile device, and when the mobile devices move from one location to the next, a new transaction manager is assigned to produce a new proxy transaction. The two-tiered transaction model is inspired by the data replication schemes, where there is a master copy of the data but for multiple replicas.  The replicas are considered to be on the mobile device but can make changes to the master copy if the connection to the wireless network is strong enough.  If the connection is not strong enough, then the changes will be made to the replicas and thus, it will show as committed on these replicas, and it will still be made visible to other transactions.

The multi-database transaction model uses asynchronous schemes, to allow a mobile user to unplug from it and still coordinate the transaction.  To use this scheme, five queues are set up: input, allocate, active, suspend and output. Nothing gets committed until all five queues have been completed. Pro-motion transactions come from nested transaction models, where some transactions are completed through fixed hosts and others are done in mobile hosts. When a mobile user is not connected to the fixed host, it will spark a command such that the transaction now needs to be completed in the mobile host.  Though carrying out this sparked command is resource-intensive.  Finally, the Toggle transaction model relies on software on a pre-determined network and can operate on several database systems, and changes made to the master database (global) can be presented different mobile systems and thus concurrency is fixed for all transactions for all databases (Panda et al, 2011).

At a cursory glance, these models seem similar but they vary strongly on how they implement the ACID properties in their transaction (see table 1) in the next section.

ACID Properties and their flaws

Jim Gray in 1970 introduced the idea of ACID transactions, which provide four guarantees: Atomicity (all or nothing transactions), Consistency (correct data transactions), Isolation (each transaction is independent of others), and Durability (transactions that survive failures) (Mapanga & Kedebu, 2013, Khachana, 2011).  ACID is used to assure reliability in the database system, due to a transaction, which changes the state of the data in the database.

This approach is perfect for small relational centralized/distributed databases, but with the demand to make mobile transactions, big data, and NoSQL, ACID may be a bit constricting.  The web has independent services connected together relationally, but really hard to maintain (Khachana, 2011).  An example of this is booking a flight for a CTU Doctoral Symposium.  One purchases a flight, but then also may need another service that is related to the flight, like ground transportation to and from the hotel, the flight database is completely different and separate from the ground transportation system, yet sites like Kayak.com provide the service of connecting these databases and providing a friendly user interface for their customers.  Kayak.com has its own mobile app as well. So taking this example further we can see how ACID, perfect for centralized databases, may not be the best for web-based services.  Another case to consider is, mobile database transactions, due to their connectivity issues and recovery plans, the models aforementioned cover some of the ACID properties (Panda et al, 2011).  This is the flaw for mobile databases, through the lens of ACID.

Model Atomicity Consistency Isolation Durability
Report & Co-transaction model Yes Yes Yes Yes
Kangaroo transaction model Maybe No No No
Two-tiered transaction model No No No No
Multi-database Transaction model No No No No
Pro-motion Model Yes Yes Yes Yes
Toggle transaction model Yes Yes Yes Yes

Table 1: A subset of the information found in Panda et al (2011) dealing with mobile database system transaction models and how they use or not use the ACID properties.

 

CAP Properties and their trade-offs

CAP stands for Consistency (just like in ACID, correct all data transactions and all users see the same data), Availability (users always have access to the data), and Partition Tolerance (splitting the database over many servers do not have a single point of failure to exist), which was developed in 2000 by Eric Brewer (Mapanga & Kadebu, 2013; Abadi, 2012).  These three properties are needed for distributed database management systems and is seen as a less strict alternative to the ACID properties by Jim Gary. Unfortunately, you can only create a distributed database system using two of the three systems so a CA, CP, or AP systems.  CP systems have a reputation of not being made available all the time, which is contrary to the fact.  Availability in a CP system is given up (or out-prioritized) when Partition Tolerance is needed. Availability in a CA system can be lost if there is a partition in the data that needs to occur (Mapanga & Kadebu, 2013). Though you can only create a system that is the best in two, that doesn’t mean you cannot add the third property in there, the restriction only talks applies to priority. In a CA system, ACID can be guaranteed alongside Availability (Abadi, 2012)

Partitions can vary per distributed database management systems due to WAN, hardware, a network configured parameters, level of redundancies, etc. (Abadi, 2012).  Partitions are rare compared to other failure events, but they must be considered.

But, the question remains for all database administrators:  Which of the three CAP properties should be prioritized above all others? Particularly if there is a distributed database management system with partitions considerations.  Abadi (2012) answers this question, for mission-critical data/applications, availability during partitions should not be sacrificed, thus consistency must fall for a while.

Amazon’s Dynamo & Riak, Facebook’s Cassandra, Yahoo’s PNUTS, and LinkedIn’s Voldemort are all examples of distributed database systems, which can be accessed on a mobile device (Abadi, 2012).  However, according to Abadi (2012), latency (similar to Accessibility) is critical to all these systems, so much so that a 100ms delay can significantly reduce an end-user’s future retention and future repeat transactions. Thus, not only for mission-critical systems but for e-commerce, is availability during partitions key.

Unfortunately, this tradeoff between Consistency and Availability arises due to data replication and depends on how it’s done.  According to Abadi (2012), there are three ways to do data replications: data updates sent to all the replicas at the same time (high consistency enforced); data updates sent to an agreed-upon location first through synchronous and asynchronous schemes (high availability enforced dependent on the scheme); and data updates sent to an arbitrary location first through synchronous and asynchronous schemes (high availability enforced dependent on the scheme).

According to Abadi (2012), PNUTS sends data updates sent to an agreed-upon location first through asynchronous schemes, which improves Availability at the cost of Consistency. Whereas, Dynamo, Cassandra, and Riak send data updates sent to an agreed-upon location first through a combination of synchronous and asynchronous schemes.  These three systems, propagate data synchronously, so a small subset of servers and the rest are done asynchronously, which can cause inconsistencies.  All of this is done in order to reduce delays to the end-user.

Going back to the Kayak.com example from the previous section, consistency in the web environment should be relaxed (Khachana et al, 2011).  Further expanding on Kayak.com, if 7 users wanted to access the services at the same time they can ask which of these properties should be relaxed or not.  One can order a flight, hotel, and car, and enforce that none is booked until all services are committed. Another person may be content with whichever car for ground transportation as long as they get the flight times and price they want. This can cause inconsistencies, information being lost, or misleading information needed for proper decision analysis, but systems must be adaptable (Khachana et al, 2011).  They must take into account the wireless signal, their mode of transferring their data, committing their data, and load-balance of incoming requests (who has priority to get a contested plane seat when there is only one left at that price).  At the end of the day, when it comes to CAP, Availability is king.  It will drive business away or attract it, thus C or P must give, in order to cater to the customer.  If I were designing this system, I would run an AP system, but conduct the partitioning when the load/demand on the database system will be small (off-peak hours), so to give the illusion of a CA system (because Consistency degradation will only be seen by fewer people).  Off-peak hours don’t exist for global companies or mobile web services, or websites, but there are times throughout the year where transaction to the database system is smaller than normal days. So, making around those days is key.  For a mobile transaction system, I would select a pro-motion transaction system that helps comply with ACID properties.  Make the updates locally on the mobile device when services are not up, and set up a queue of other transactions in order, waiting to be committed once wireless service has been restored or a stronger signal is sought.

References

  • Abadi, D. J. (2012). Consistency tradeoffs in modern distributed database system design: CAP is only part of the story. IEEE Computer Society, (2), 37-42.
  • Khachana, R. T., James, A., & Iqbal, R. (2011). Relaxation of ACID properties in AuTrA, The adaptive user-defined transaction relaxing approach. Future Generation Computer Systems, 27(1), 58-66.
  • Mapanga, I., & Kadebu, P. (2013). Database Management Systems: A NoSQL Analysis. International Journal of Modern Communication Technologies & Research (IJMCTR), 1, 12-18.
  • Panda, P. K., Swain, S., & Pattnaik, P. K. (2011). Review of some transaction models used in mobile databases. International Journal of Instrumentation, Control & Automation (IJICA), 1(1), 99-104.

Adv Topics: MapReduce and Incremental Computation

Data usually gets update on a regular basis. Connolly and Begg (2014) defined that data can be updated incrementally, only small sections of the data, or can be updated completely. An example of data that can be updated incrementally are webpages, computer codes, stale data, data-at-rest, bodies of knowledge, etc. Whereas, some examples of data that can be updated completely are: weather data, space weather data, social media data, data-in-motion, dynamic data, etc. Both sets of data provide their own unique challenges when it comes to data processing. On average, analyzing web data, new to old data can range from 10-1000x (Sakr, 2014). Thus, the focus of this discussion is on incremental data update and how to process data in between two data processing runs.

Incoop is an extension of Hadoop to allow for processing incremental changes on big data, by splitting the main computation to its sub-computation, logging in data updates in a memoization server, while checking the inputs of the input data to each sub-computation (Bhatotia et al., 2011; Sakr, 2014). These sub-computations are usually mappers and reducers (Sakr, 2014). Incremental mappers check against the memoization servers, and if the data has already been processed and unchanged it will not reprocess the data, and a similar process for incremental reducers that check for changed mapper outputs (Bhatotia et al., 2011).

Subsequently, MapReduce is an analytical engine and pattern that takes advantage of distributed systems while keeping the processes and data in one machine (Sadalage & Fowler, 2012). There are a few key principles to using the MapReduce framework and Hadoop efficiently to improve incremental computation:

  • Data partitioning: The MapReduce framework aids in partitioning the data into similar size sets into Hadoop Distributed File System, aka HDFS (Lublinsky, Smith, & Yakubovich, 2013). Thus, MapReduce can support smaller sets of data stored in HDFS. This is part of the scalability of the cluster.
  • Fault tolerance and durability: Given that data can be partitioned to tiny chunks across thousands of computations nodes and run in parallel sometimes these nodes can fail (Sakr, 2014). The MapReduce framework replicates the data in the background and can launch backup jobs if a node fails (Lublinsky et al., 2013; Sakr, 2014). Thus, failure doesn’t disrupt the data processing. However it does increase the number of processors needed (Connolly & Begg, 2014).
  • Parallelization: The partitioned input data are considered as independent sets of data, such that the mapper functions can process the data in a parallel environment (Lublinsky et al., 2013; Sakr, 2014). This principle allows for the sub-functions within the mapper and reducer function to handle smaller data. It allows a data analyst to focus on the main problem rather than low-level parallel coding abstraction, like multithreading, file allocation, memory management, etc. (Sakr, 2014). Serialization does not allow for small incremental updates for large data (Connolly & Begg, 2014).
  • Data reuse: There is no need to read or write of intermediate data, thus preserving the input data to enable the data to be reused because it is unchanged (Lublinsky et al., 2013; Sakr, 2014).
  • Self-Adjusting Computation: Used for incremental computation, which only allows mappers and reducers only work on the smaller size sets of data that are impacted by the change (Sakr, 2014).

Both Bhatotia et al. (2011) and Sakr (2014), suggested an Inc-HDFS which is also an extension of the HDFS, for partitioning data based on content and removal of data duplication. There is the limitation of this approach where the number of files may be grouped in too many content bins or too little content bins and thus may not be evenly be spaced out (Bhatotia et al., 2011). Thus, invoking: too many mapper functions can create an infrastructure overhead, which increases resources and thus cost, or too few mapper functions can create huge workloads for certain types of computational nodes, or too many reducers can provide too many outputs, and too little reducers can provide too little outputs (Lublinsky et al., 2013; Sakr, 2014). A constraint must be added on both ends of the spectrum to allow for evenly distributed data sets (Bhatotia et al., 2011).

Subsequently, the Hadoop out-of-the-box product scheduler doesn’t account for memoization server data, therefore is not built for incremental analysis. Thus, Incoop has a memoization-aware scheduler, that schedules the sub-computations based on the affinity of the task and allows for efficient use of previously (Bhatotia et al., 2011; Sakr, 2014). The scheduler can run tasks on computational nodes that are either faster or locally to where the data is stored (Bhatotia et al., 2011). Using this type of scheduler, the scheduler should place priority on whether it is faster to conduct a data movement and run the task at a faster computational node or reduce data movement, and processes the data locally, while still making effective use of unchanged and processed data.

In the end, a practical application of this technique when it comes to analyzing web data would be to first partition the web data by its content. Scientific content can go in one context partition, corporate financial content into another context partition, etc. under the Inc-HDFS framework. These partitions are capped in size to allow for proper load balance. Incoop will then run the MapReduce function to process the data distributively through using parallel processes. If the data gets updated, like a new corporate update to the SEC 10K data, it will be recorded by the memoization server. This will allow for Incoop to be able to process that incremental change from the corporate financial content partition because it was using the memoization-aware scheduler, and reprocess the data through mapping and reducing function on just this small and partitioned dataset. Therefore, making effective use of unchanged and processed data.

Resources:

  • Sadalage, P. J., Fowler, M. (2012). NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence, (1st ed.). Vitalbook file.
  • Sakr, S. (2014). Large Scale and Big Data, (1st ed.). Vitalbook file.ok