Technical blog by Matt Kubej

Graphs with Spidey DB

April 9th, 2021

A recent challenge and exploration I engaged in involved building an in-memory graph database in C. I sought inspiration from Redis, but I have little experience in C outside of academia and little experience with graph databases. So, this served itself as a great learning experience to build something of medium complexity in unfamiliar territory. Additionally, I have always been impressed by the visualizations that D3 has allowed people to construct, so I desired to give that a swing at illustrating my graph. Lastly, SvelteKit just had a beta release, so I thought I could learn Svelte while I was at it to provide the user interface and serve D3.

Illustrated below is a small demo of what I pulled together.

demo

Now to go into how I overengineered this for educational purposes and fun.

Building the graph

The journey began by constructing a graph data structure within C. I desired to replicate a Redis-like experience, which allowed for speedy sets and gets. Additionally, I wanted to be able to find the nearest neighbors of a vertex quickly. With these considerations in mind, I settled on an adjacency list. This allowed for constant time lookups of vertices. As for gathering neighbors, this would take O(v+e) time, where v is the number of neighbors and e is the number of edges between the root vertex and neighboring vertices.

An adjacency list is effectively a hash map of linked lists. Each entry within the hash map would represent a vertex and the entry would contain a linked list, which represents the edges coming from the vertex and to directly neighboring vertices. Additionally, these entries could contain the data associated with the vertex.

Example graph:
(a) <---> (b) <---> (c)

Adjacency list representation:
[a]: (b) -->
[b]: (a) --> (c) -->
[c]: (c) -->

In the example above, the graph has three vertices (a, b, and c) along with four edges, which consists of a to b, b to a, b to c, and c to b. The adjaceny list above illustrautes what a hash map might look like to represent that particular graph. It would have an entry for each vertex, where the key is represented by square brackets. The value associated with each entry is a linked list of the edges stemming off of that vertex. The ordering of the edges does not matter in my circumstance, because I just want to be able to add and retrieve them quickly. Thus, in favor of performance, to add an edge to a list consists of placing it on the head, so I do not need to forgo iterating across the list to append to it.

Retrieving a vertex from the adjacency list can be done in constant time due to the utilization of a hash map data structure. To find the neighboring edges of a vertex involves iterating through the linked list of edges associated with that vertex, which results in O(e) time, where e is the number of edges.

Due to my decision to use C for educational purposes, this required me to build my own hash map data structure to support my adjacency, which is a luxury experienced with the standard libraries of many other programming languages. I came across a great write up on building hash tables in C called C/HashTables and effectively replicated the hash map it outlined.

Networking in C

With some success in representing my graph, I moved onto the next piece of the puzzle, which is how do I communicate with my graph database? Due to my previous experience with RESP and the simplicitly it offered for communication, I selected this as the communication protocol to use over TCP. An additional benefit of this decision meant I could use the redis-cli packaged with Redis as my client for communicating with my graph database, which meant I did not have to bother with constructing a client during development to communicate with my database. This served as a large time saver.

I followed the same pattern and strategy as outlined by Redis' RESP documentation, but I invented my own commands to instruct my graph database. Listed below outlines the commands I intended to support. With these in combination, I could construct a graph and search for neighboring vertices of a vertex.

commanddescription
GETVERTEXretrieve the value of a vertex
SETVERTEXset the value of a vertex
DELVERTEXdelete a vertex and associated edges
SETEDGEset an edge, linking two vertices
DELEDGEdelete an edge, linking two vertices
GETGRAPHretreive the entire graph
GETNEIGHBORSretreive the neighbors a vertex

For awareness, I did introduce the COMMAND command as a means to please redis-cli. It invokes this command to Redis on start-up to learn what commands are available and it will block the start-up until the message is received. So, by introducing this command, it allowed me to connect to my database without issue.

With a specification for communication and understanding of how I would manage my graph, I need to be able to accept and handle client communications over the network. This required me to roll up my sleeves a little bit and learn socket programming in C. One guide I found to be a staple and frequently found in academia is Beej's Guide to Network Programming, which it proved to be extremely useful. I will not go into detail of socket programming here, but I leaned heavily on this guide to accept incoming connections, manage multiple clients, process requests, and serve responses. In terms of the codebase, this is likely the most complex portion of the implementation. If you would like to learn more, then I would recommend reading Beej's guide.

Listed below is an example output of some communication with the graph database via redis-cli:

❯ redis-cli -p 6425
127.0.0.1:6425> SETVERTEX a-key a-value
OK
127.0.0.1:6425> SETVERTEX b-key b-value
OK
127.0.0.1:6425> SETEDGE a-key b-key
OK
127.0.0.1:6425> GETGRAPH
1) 1) "a-key"
   2) "b-key"
2) 1) 1) "a-key"
      2) "b-key"
   2) 1) "b-key"
      2) "a-key"
127.0.0.1:6425>

The output above creates a vertex with a key of a-key and a value of a-value and a vertex with a key of b-key and a value of b-value. The vertices are joined together with an edge and then the graph is returned. The graph is represented with two arrays. The first array is a list of the vertices. The second array is a list of the edges.

Searching for Neighbors

The last piece to round out the graph database implementation involved the ability to search for nearby neighbors of a vertex. The algorithm I selected to accomplish this involved using breadth first search with a queue. The following list details in the algorithm in short.

  1. Enqueue the root vertex into an unvisited queue
  2. Dequeue an unvisited vertex
  3. Mark the dequeued vertex as visited
  4. Iterate through the edges of the dequeued vertex and enqueue each neighboring vertex in a queue
  5. Repeat steps 2 through 4 until the unvisted queue is empty

During the execution of this algorithim, I saved the processed vertices and edges to represent the graph subset describing those neighboring the root vertex. After completion, I could then convert the graph to a RESP representation to communicate the result back to the requesting client.

Listed below is an example output of the GETNEIGHBORS:

❯ redis-cli -p 6425
127.0.0.1:6425> SETVERTEX a 1
OK
127.0.0.1:6425> SETVERTEX b 2
OK
127.0.0.1:6425> SETVERTEX c 3
OK
127.0.0.1:6425> SETEDGE a b
OK
127.0.0.1:6425> SETEDGE b c
OK
127.0.0.1:6425> GETNEIGHBORS a 1
1) 1) "a"
   2) "b"
2) 1) 1) "a"
      2) "b"
127.0.0.1:6425>

The output above creates a graph that would look as follows.

(a) <---> (b) <---> (c)

The GETNEIGHBORS command receives the root vertex as the first argument and the distance to search as the second argument. The example specifices a distance of one for vertex a, so it additionally returns b and the associated edge that connects them. However, it does not return c as that vertex is two edges away.

Visualizing the Graph

Now that I had roughly working graph I wanted to see if I could put it to use by having a consuming client illustrate the graph with a visualization. Due to wanting to explore SvelteKit and D3, this meant communicating with my graph database via Node and TypeScript. I previously wrote a basic Redis clone with TypeScript called redis-ts, which I could repurpose some of that code for connecting to my server, producing RESP, and processing RESP. This allowed me to build a simple API within TypeScript for talking with my database. The implementation can be found here. However, if you're curious about the Redis clone, then I'd recommend taking a look at my previous blog post called Redis clone in TypeScript.

With this in place, I could programmatically build, retrieve, and search my graph via TypeScript. However, I wanted to be able to see this within my browser. I set out to come up to speed on SvelteKit, which had a beta release on March 23rd, 2021. I have previous experience with other utilities such as React for building user interfaces, but have been hearing more about Svelte and wanted to form my own opinion of the technology, so I dove right in.

SvelteKit and Svelte take a different approach than other popular frameworks at the moment by not utilizing a virtual DOM by leveraging compilation and focused on eliminating boilerplate code through minimal syntax. This allowed me to quickly construct self-contained and composable components. Additionally, due to their syntactic decision making and design, Svelte allowed me to visually present my graph on a webpage with little code. Overall, it provided very little barrier of entry to produce satisfying results.

However, SvelteKit is in beta and I did hit a number of snags or oddities along the way, which I would anticipate to be resolved in future releases. For example, I found nesting folders and files within the src/lib folder to varying degrees produced transpilation errors or timing issues with the loading of my CSS. The most recent one I encountered involved challenges with building production assets using their provided build script. I hope to find opportunities to contribute back in this space even though I am new to this solution, but I was able to have a lot of succes with it considering it is in beta.

As for the D3 portion of the project, I referenced examples on Observable and documentation on the D3 GitHub, then retrofitted the examples to be more Svelte-esque. D3 produces great results, but the guidance and documentation for constructing these visualizations seems rather scattered and challenging to navigate. This required a lot of trial and error while fumbling around the documentation to piece together the various APIs. Similarly, there are not many examples of D3 in Svelte at the moment, but I did come across an interesting posting which pushed me in a good direction to employ Svelte's expression language in accordance with D3 rather than utilizing DOM manipulation via JavaScript. The write up on this is called Animated, Responsive, and Reactive Data Visualization with Svelte .

Concluding Thoughts

Building this project allowed me the opportunity to explore some less familiar areas and problems I less frequently encounter due to the selection of C. This provided me practice at building data structures like hash maps, queues, linked lists, and graphs, while trying to optimize them for my use case. Unless I am really squeezing for performance or I am technically restricted in some fashion, then I think I am rather unlikely to select C for many of my personal projects going forward. It gives you a lot of control, but the overhead of building your own data structures and memory management becomes quite cumbersome quickly.

As for SvelteKit, I think it has the opportunity to present a lot of value due to it's low barrier of entry, but it needs more time to grow out of beta. If I were to build the same user experience with something like React, I could likely build something more production-ready, but it would require more code. For simple web applications, SvelteKit offers a lot of promise once it works out the kinks.

You can find Spidey DB at: https://github.com/kubejm/spidey-db

You can find the Spidey DB demo at: https://github.com/kubejm/spidey-db-demo