Conventional wisdom says that concurrent architectures are bound by tradeoffs between efficiency and complexity. More concurrency offers more efficiency, which comes at the cost of architectural complexity. Implementing concurrency requires more cognitive overhead. The code may be much more fragmented, making things like debugging and maintenance tough. More importantly, concurrency can also compromise consistency, because concurrent operations (e.g. near-simultaneous read and writes) can lead to data integrity issues. Consumers and developers alike have some expectation of immutability. In this post I talk through my notes on reading Pat Helland’s “Immutability Changes Everything”.
What is Concurrency?
In my post on the Actor Model, I briefly touched on concurrency, which means a program that has been broken down into steps that can be successfully executed independently and even out-of-order. Concurrent architectures, when implemented well, offer tremendous speed, scaling, and efficiency (e.g. time, power usage). In some cases, they may also offer security advantages (e.g. handling DDOS attacks). Concurrency can be especially useful for I/O bound tasks, or long running programs that require interaction with lots of servers or users.
What is Immutability?
Probably like most programmers, when I think about immutability I’m generally thinking at a higher level of abstraction; e.g. objects that can and can’t change. In general, primitives tend to be immutable while container-like types tend to be mutable, though this isn’t a hard and fast rule in every language. For instance, in Python, the primitives
bool are immutable, while
set are all mutable. On the other hand, Python
tuples (non-primitive) are immutable. On the other hand, in Go data structures are not immutable, except for strings.
Of course in the sense that Helland is using in his piece, immutability means that data inside a transactional database does not change — except to accumulate.
“We can now afford to keep immutable copies of lots of data, and one payoff is reduced coordination challenges.” - Pat Helland
In other words, coordination is now a lot more expensive than keeping more copies and adopting the practice of not deleting things.
I wish this were how people actually used RDBMS in practice! It would certainly make the kind of analytics I do a bit more straightforward. But from an infrastructure perspective, I can see how this is a robust strategy for distributed databases, particularly as the velocity of transactions increases. Using this append-only strategy, discrepancies between copies of the database can be easily rectified via updates, which will only ever come in the form of new writes, never re-writes or deletions. Maintaining additional redundant copies of the database can also increase our confidence level.
One of the things that stood out to me most about Helland’s piece is that he calls himself a “database old timer” in the sense that he came up thinking of RDBMS as the only type of database. And yet he takes a fairly open-minded perspective towards the “DataSet” as he calls it. He defines a “DataSet” as a snapshot or extract from a database; it’s conceived as a fixed and immutable set of tables, including both schema information and metadata (e.g. version) as well as the data. However, I take this to mean a “DataSet” can really be any downstream database view, including one that might get used to populate something like a Neo4J graph or ElasticSearch index.
Helland still frames relational databases as the source of truth (the “Data on the Inside”), casting these secondary data products (reports, flat files, NoSQL stores) as “Data on the Outside”. Thinking of these data (or DBs) as “outside” allows us to treat them with different expectations than we would of a traditional transactional database; because they’re understood to be immutable, they take a lot less work in terms of management (e.g. we don’t have to worry about locking or controlling updates).
His view seems somewhat different from the tack that other traditionalists in the community took during the ‘NoSQL revolution’.
“By watching and monitoring the read usage of a DataSet, you may realize new optimizations (e.g. new indices) are possible.” - Pat Helland
In particular, the above line struck me. While we may accept as best practice that downstream data products are immutable, and thus should not directly impact the original data store (though I’ve observed this more than once!), they are nonetheless informed by the analytic needs of downstream consumers. As such these data products can help anticipate large-scale changes in transactional patterns happening in the underlying RDBMS, potentially even informing the design, architecture, and deployment of RDBMS, or potentially of their underlying server communication or log update protocols.
The Truth is the Log
“The truth is the log. The database is a cache of a subset of the log.” - Pat Helland
Append-only computing describes a general approach and comprises multiple strategies, including ones that seem like they may have outlived their usefulness for a lot of modern applications (like single master computing, which has been superseded by distributed single master like Paxos).
Vocab & Acronyms
- Dynamic random-access memory (DRAM) is a type of random access semiconductor memory that stores each bit of data in separate, small capacitors within an integrated circuit.
- Copy-on-write (COW) (aka implicit sharing or shadowing) is a technique for coordinating modifications in a memory-efficient way. For example, let’s say we log into a shared virtual machine and prepare to make some changes. Under COW, the kernel intercepts to create a copy or fork, but not until we actually start making changes. If we don’t end up making any changes, the copy never gets made.
- A solid-state drive (or solid-state disk) (SSD) is a storage device that uses integrated circuit assemblies as memory to store data persistently.
- A hard disk drive (aka hard disk, hard drive, or fixed disk) (HDD) is an electromechanical data storage device that stores and retrieves information using hard spinning disks coated in magnetic material.
- A log-structured file system (LSF) is a file system in which data and metadata are written sequentially to a circular buffer called a log.
- A log-structured merge tree (LSM) is a data structure that maintains data in key-value pairs, which provides indexed access to files with high insert volume (e.g. transactional log data).
- A database management system (DBMS) is a software application that interacts with end users, other applications, and the database itself to capture and analyze data.
- A “DataSet” is a fixed and immutable set of tables, a snapshot or extract from the database. It includes both schema information and data. It is created, consumed for reading, then deleted. DataSets may be relational or represented as a graph, a hierarchy (JSON), etc. Because it is immutable, there’s no need for locking and no worries about controlling updates. It doesn’t take any management.
- A linear version history is strongly consistent. One version replaces another. There’s one parent and one child. Each version is immutable. Each version has an identity. Typically, each version is viewed as a replacement for the earlier versions.
- A directed acyclic graph of version history may have many parents and/or many children. This is sometimes called eventual consistency.
- Strongly consistent (ACID) transactions appear as if they run in a serial order. This is sometimes called serializability. ACID stands for:
- Atomicity: An atomic transaction is an indivisible and irreducible series of database operations such that either all occur, or nothing occurs.
- Consistency: Any given database transaction must change affected data only in valid ways, i.e. according to all defined rules individually and in combination.
- Isolation: Determines how transaction integrity is visible to other users and systems. Lower isolation levels increase the ability of many users to access the same data at the same time, but increase the number of concurrency effects (such as dirty reads or lost updates) users might encounter. A higher isolation level reduces the types of concurrency effects, but requires more system resources and increases the chances that one transaction will block another.
- Durability: guarantees that transactions that have committed will survive permanently. Durability can be achieved by committing log records to non-volatile storage before considering the transaction complete. In distributed transactions, all servers must coordinate before a commit can be acknowledged. This is usually done by a two-phase commit protocol.
- Google File System (GFS) is a proprietary distributed file system developed by Google to provide efficient, reliable access to data using large clusters of commodity hardware.
- Hadoop Distributed File System (HDFS) is a distributed file-system that stores data on commodity machines, providing very high aggregate bandwidth across the cluster.
- A Globally Unique Identifier is a 128-bit number used to identify information in computer systems.
- A Service Level Agreement is a commitment between a service provider and a client regarding things like quality, availability, and responsibilities.
- Wear leveling is a form of copy-on-write and treats each version of an SSD block as an immutable version.
- Shingled Disk Systems aim to increase storage density and overall per-drive storage capacity via shingled magnetic recording.
- Bernstein, P.; Hadzilacos, V.; Goodman, N. (1987). “Concurrency Control and Recovery in Database Systems”, Addison Wesley, ISBN 0-201-10715-5.
- Dean, J.; Ghemawat, S. (2004). “MapReduce; Simplified Data Processing on Large Clusters”. OSDI ’04: 6th Symposium on Operating System Design & Implementation.
- DeCandia, G.; Hastorun, D.; Jampani, M.; Kakulapati, G. Lakshman, A.; Pilchin, A.; Sivasubramanian, S.; Vosshall, P. Vogels, W. (2007). “Dynamo: Amazon’s Highly Available Key-Value Store”. Proc of the 21st ACM Symp on Operating Systems Principles.
- Ghemawat, S.; Gobioff, H.; Leung, S. (2003) “The Google File System”. Proceeedings of the 19th ACM Symposium on Operating Systems Principles – SOSP ‘03
- Gibson, G.; Ganger, G. (2011) “Principles of Operation for Shingled Disk Devices”. Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-11-107.
- Helland, P. (2005) “Data on the Outside versus Data on the Inside” Proceedings of the 2005 CIDR Conference (Conference on Innovative Database Research).
- Helland, P. (2014) “Heisenberg Was on the Write Track”. Abstract: Proceedings of the 2015 CIDR Conference (Conference on Innovative Database Research).
- Isard, M.; Budiu, M.; Yu, Y.; Birrell, A.; Fetterly, D. (2007) “Dryad: Distributed Data-Parallel Programs from Sequential Building Blocks” European Conf on Computer Systems (EuroSys).
- Karger, D.; Lehman, E.; Leighton, T.; Panigraphy, R.; Levine, M.; Lewin, D. (1997). “Consistent Hashing and Random Trees: Distributed Caching Protocols for Relieving Hot Spots on the World Wide Web”. Proc. of the 29th Annual ACM Symp on Theory of Computing.
- Lamport, L. (1998). “The Part-Time Parliament”, ACM Transactions on Computer Systems (TOCS), Volume 16, Issue 2, May 1998.
- McKusick, M.; Quinlan, S.; “GFS: Evolution on Fast Forward” (2009) ACM Queue, August 7, 2009.
- O’Neil, P; Cheng, E.; Gawlick, D.; O’Neil, E. (1996) “The Log-Structured Merge-Tree (LSM-tree)”. Acta Informatica 33 (4).
- Rosenblum, M.; Ousterhout, J. (1992) “The Design and Implementation of a Log-Structured File System”. ACM Transactions on Computer Systems, Vol. 10, Issue 1.