Category: Architecture

Jessica Kerr discusses methods for determining data lineage, particularly in distributed systems:

Can you take a piece of data in your system and say what version of code put it in there, based on what messages from other systems? and what information a human viewed before triggering an action?

Me neither.

Why is this acceptable? (because we’re used to it.)
We could make this possible. We could trace the provenance of data. And at the same time, mostly-solve one of the challenges of distributed systems.

This is an interesting essay; read the whole thing.

Lukas Eder discusses when to use a relational database versus some non-relational database:

This question obviously assumes that you’re starting out with an RDBMS, which is classically the database system that solves pretty much any problem decently enough not to be replaced easily. What does this mean? Simply put:

• RDBMS have been around forever, so they have a huge advantage compared to “newcomers” in the market, who don’t have all the excellent tooling, community, support, maturity yet
• E.F. Codd’s work may have been the single biggest influence on our whole industry. There has hardly been anything as revolutionary as the relational model ever since. It’s hard for an alternative database to be equally universal, i.e. they’re mostly solving niche problems

Having said so, sometimes you do have a niche problem. For instance a graph database problem. In fact, a graph is nothing fundamentally different from what you can represent in the relational model. It is easy to model a graph with a many-to-many relationship table.

If you want a checklist, here’s how I would approach this question (ceteris paribus and limiting myself to about 100 words):

1. Are you dealing with streaming millions of rows per second, or streaming from tens of thousands of endpoints concurrently?  Kafka and the Hadoop streaming stack.
2. Is your problem something that you’ve already solved with a relational database, and your solution works well enough?  Relational database.
3. Are there multiple “paths” to get to interesting data?  Relational database.
4. Shopping carts (write-heavy, focused on availability over consistency)?  Riak/Cassandra/Dynamo at large scale, else relational database.
5. Type duplication?  Relational database.
6. Petabytes of data being analyzed asynchronously?  Hadoop.
7. Other data platforms if they fit specific niche requirements around data structure.

There’s a lot more to this discussion than a simple numbered list, but I think it’s reasonable to start with relational databases and move away if and only if there’s a compelling reason.

Erik Darling describes partitioned views:

Hooray. Now you have to type less.

Partitioned views don’t need a scheme or a function, there’s no fancy syntax to swap data in or out, and there’s far less complexity in figuring out RIGHT vs LEFT boundaries, and leaving empty partitions, etc. and so forth. You’re welcome.

A lot gets made out of partition level statistics and maintenance being available to table partitioning. That stuff is pretty much automatic for partitioned views, because you have no choice. It’s separate tables all the way down.

Partitioned views, AKA SQL Server 2000 partitioning.  I think my favorite use case for them today is to serve as a combination of hot data in a memory-optimized table and cold data on disk.

Diana Tkachenko describes a pattern for reducing “prod doesn’t look like stage” types of errors:

Immutable server pattern makes use of disposable components for everything that makes up an application that is not data. This means that once the application is deployed, nothing changes on the server – no scripts are run on it, no configuration is done on it. The packaged code and any deploy scripts is essentially baked into the server. No outside process is able to modify the contents after the server has been deployed. For example, if you were using Docker containers to deploy your code, everything the application needs would be in the Docker image, which you then use to create and run a container. You cannot modify the image once it’s been created, and if any changes do need to take place, you would create a new image and work with that one.

In our case, we use AWS Amazon Machine Images (AMIs) to accomplish the same thing. We make heavy use of Amazon Linux machines, which are Redhat-based, and thus package the code into RPMs^[2]. The RPMs define all the dependencies for running the application, the code itself, and any startup scripts to run on bootup. The RPM is then installed on a clean base image of Amazon Linux, and an image is taken, resulting in an AMI. This AMI is synonymous with “immutable server” – it cannot be changed once it is created. The AMI is then deployed into an Auto Scaling Group(ASG) and attached to the Elastic Load Balancer (ELB). In this post, I’ll guide you through for a closer look at every step of this Immutable Server deploy pipeline. I’ll then go into how and why we embedded planned failures into this system. At the end, I’ll share the insights we’ve gained into the pros and cons of deploying in this way.

This is a very interesting concept.  I’ve heard of no-patch servers (where, instead of patching live servers, you spin up a new VM with the operating system updates and spin down the old one), but this takes the idea one step further.

Kennie Pontoppidan discusses why you don’t want your transactional system on the same instance as your warehouse:

Mixing OLTP and data warehouse workload types in the same SQL Server instance is not a good idea. Why? Well, because the workload types don’t mix. As described above, an ideal buffer cache for an OLTP system contains pages used in searches as well as pages for the most used customer/items etc. Running a few data warehouse queries will change the buffer cache to contain (mostly) all dimension data. This will make OLTP operations slow. On the other hand, the ideal buffer cache in a data warehouse will contain all dimension data, which makes aggregations fast because all dimension lookups will be done in memory. OLTP operations might flush out parts of this dimension data from the cache, hurting the performance for data warehouse queries. The two workload types will keep on fighting about the content of the buffer cache, making both systems suffer.

Except in very small systems (where it doesn’t much matter), this is sound advice.

Dan Luu has a great article from a couple years ago on when a random cache eviction policy might be preferable to Least Recently Used:

Once upon a time, my computer architecture professor mentioned that using a random eviction policy for caches really isn’t so bad. That random eviction isn’t bad can be surprising – if your cache fills up and you have to get rid of something, choosing the least recently used (LRU) is an obvious choice, since you’re more likely to use something if you’ve used it recently. If you have a tight loop, LRU is going to be perfect as long as the loop fits in cache, but it’s going to cause a miss every time if the loop doesn’t fit. A random eviction policy degrades gracefully as the loop gets too big.

In practice, on real workloads, random tends to do worse than other algorithms. But what if we take two random choices and just use LRU between those two choices?

Here are the relative miss rates we get for SPEC CPU¹ with a Sandy Bridge-like cache (8-way associative, 64k, 256k, and 2MB L1, L2, and L3 caches, respectively). These are ratios (algorithm miss rate : random miss rate); lower is better. Each cache uses the same policy at all levels of the cache.

Dan writes at a depth I appreciate and on topics I often don’t understand (particularly when he gets into CPU engineering details).

Dave Copeland discusses over-engineering problems:

The main problem with an over-engineered solution is that it takes longer to ship than is necessary. By definition, we are doing more than is necessary, and that will take longer to ship. There’s almost never a reason to prefer longer ship-times over shorter ones, all things being equal.

The more serious problem with over-engineering is the carry cost.

A carrying cost is a cost the team bears for having to maintain software and infrastructure. Each feature requires tests, monitoring, and maintenance. Each new feature is made in the context of those that came before it. This is why a feature that might’ve taken one week when the project was new requires a month to make in more mature project.

Read the whole thing and simplify your solutions.

Kyle Kingsbury explains that sequential, serializable, and strictly serializable consistency models cannot provide locality:

We often speak of locality as a property of subhistories for a particular object x: “H|x is strictly serializable, but H is not”. This is a strange thing to say indeed, because the transactions in H may not meaningfully exist in H|x. What does it mean to run [(A enq y 1) (A enq x 1)] on x alone? If we restrict ourselves to those transactions that doapply to a single object, we find that those transactions still serialize in the full history.

So in a sense, locality is about the scope of legal operations. If we take single enqueue and dequeue operations over two queues x and y, the space of operations on the composite system of x and y is just the union of operations on x and those on y. Linearizability can also encompass transactions, so long as they are restricted to a single system at a time. Our single-key, multi-operation transactions still satisfied strict serializability even in the composite system. However, the space of transactions on a composite system is more than the union of transactions on each system independently. It’s their product.

Here’s the part where I pretend that of course I understand what Kyle wrote…  Seriously, though, this is a very interesting read.