I love reading books. My wife loves reading books. We enjoy shopping for books and we live a ten minute commute away from a huge used books store. That means we have a lot of books. Like, really a lot. A little more than one thousand.
We have lost count how many times we've bought a book that we already owned. Even more often we had foregone buying a book we liked because we were sure we have already bought it - only to find out we've been mistaken and have only considered buying that very book earlier. This has become a problem. The Library Problem.
We needed a way to catalog all our books. The catalog had to be accessible from mobile devices (to look up a book while at the book store) and to be easy to use. That is to add and edit book information, of which we've needed plenty: in addition to standard set of author, title and publishing year we wanted to be able to track book series and keep the list of missing books to look out for the next time.
I admit that my research into the subject matter was not scientifically thorough. I've dug up several comparisons of existing tools and have read several blog posts of people who have faced the same problem before. I particularly recommend this one. And I have decided that no pre-existing tool will meet our growing expectations.
I have never developed an application for any end user other than myself and I didn't even know I was about to start developing one.
The initial idea was that of a "document". It had to contain some essential information about every book we own and (optionally) a second list of books we want to acquire. Excel spreadsheet seemed like a natural fit. We are both spending a better half of each workday juggling spreadsheets in Excel, so developing and maintaining such "document" appeared doable.
I have started drafting the column structure, applied some formatting tweaks and the skeleton of the "document" was ready. We have entered the test batch of books and were about to begin testing.
Obstacles arose when we were entering information for the first fifteen books. Manually typing in all the fields is tiresome and slow. Errors inevitably happen. What if I miss one letter in the author's name? That book will be lost when applying filter on author column. We need autocompletion! What if I accidentally switch the order of digits in ISBN? We won't be able to do a web lookup for that book later. We have to write a custom ISBN validator in VBA!
And the spreadsheet began to amass VBA code, data validation and conditional formatting rules. Full-blown spaghetti style. Version control has become a problem. What's the difference between versions 0.0.13 and 0.0.19? I had only a vague idea.
I stopped myself when I was about to sketch up a UserForm for data input. Excel road was leading me nowhere. It was difficult and even if all the difficulties were to be overcome it imposed some suboptimal compromises on us:
- Single table storage limited the data structures we would be able to enter and view. If the book was written by two authors, which one should come first in the "author" string? How would we do column filtering in that case? What about sort order?
- The local nature of storage meant there had to be one designated place for making changes (home laptop). Any changes made in other locations (smartphone, thumb drive) had to be agreed to be declared discardable.
- The spreadsheet had to be exported to HTML to be accessible from smartphones. XML and XSLT made this possible but not very pleasant. Although, I am rather proud of the VBA code I wrote to save/load XML data automatically upon opening the workbook. The data was completely decoupled from the representation.
I'm glad I did not waste more time pursuing this path, but it was still hard to let go. It took quite some time for me to return to this project afterwards.
A relational database was the logical solution to spreadsheet limitations. Store authors separate from books and manage how the former relate to the latter! Scientists who pioneered the relational model in 1970s were pure geniuses and now the whole world relies on their work.
I drafted the database schema on a piece of paper and have discussed it extensively with my wife. That's probably the point where book reviews were added to the requirements list. The database idea made me very enthusiastic and had swallowed a lot of free time, but that idea alone could not provide a complete solution for the problem at hand.
Data input and representation are what the Library project was all about. Yet relational database management systems provide only the storage solution, not the full package (if you don't take Microsoft Access into account). So I had to figure out how to implement the user interface on my own. I chose to write it in Python because I was somewhat familiar with the language and I enjoy how clean and readable Python code is. I'm not a programmer, so other options were either learning a new language or choosing between VBA and Bash, none of which could be considered enjoyable.
I have briefly considered building a conventional desktop GUI. The UserForm experience was still fresh and rather traumatic, so I was reluctant to dive into another UI toolkit. Even if I were to, which one should I have chosen? Qt seemed nice, but its Python support reports were contradictory. GTK? Installing it was rather tricky in Windows. Something non-crossplatform? And what about my Linux laptop?
I've had a bunch of HTML templates left from XML/XSLT operations and they could be trivially transformed to be used in conjunction with the database. While the catalog pages could be statically generated, the data input required some sort of server to interact with. And I've had zero experience with that.
Quick Google search has introduced me to the concept of web frameworks and I have semi-randomly chosen Bottle. At that moment I've had no intention to expose it to the open network, the app and the database were to be stored on the USB stick and launched locally when needed. Smartphone interaction was planned either via LAN or using saved HTML pages for read-only access when not at home. Bottle uses no dependencies other than standard library and the whole framework is packaged into a single file. It was perfect for the portable app scenario.
After the development started and I saw how difficult it is to make a web application, I've decided there is no point to confine the result of all that labor to a single computer. Why use static (maybe even outdated) HTML dumps on the smartphone when we could access full functionality of the application via web site?
The development was already going on, powered by Python and Bottle. What I was missing was the server to host my application at. Turns out launching a web site can be very cheap these days! Even as cheap as free.
I've started with a free domain name from Freenom and a free hosting from Red Hat's OpenShift. Both of them have had upsides and downsides but no downsides were significant enough to turn down the price of free. Freenom domains are considered low quality from SEO standpoint, but that was completely irrelevant in my case. OpenShift server had to receive at least one http request every 24 hours to stay awake, and I've cheated that with a cron job on my router. No inconveniences whatsoever.
I have to digress to emphasize: OpenShift was great! It was easy to set up and very convenient to use. Its documentation was very thorough and up to date. OpenShift has introduced me to some concepts I would have not encountered otherwise, like automated deploying after git push and handling proper wsgi server. And all of that was for free. I'm very thankful to Red Hat for that.
Unfortunately, as Mr. Heinlein used to reiterate, there ain't no such thing as a free lunch. Red Hat could not pay the bills for all the computer enthusiasts forever and version 3 of OpenShift imposed much tighter restrictions on the free accounts. I've had to move to a cheap virtual private server and to learn to set up and maintain an Apache instance on my own. New VPS was sometimes too slow compared to what OpenShift has offered and I might need to migrate to another hosting provider later, but everything was fine for now.
I could concentrate on the development.
Little did I know that while I was keeping myself busy with seemingly important problems such as whether to store images as blobs in database or as files on the file system a number of much more important problems were creeping up behind me. Donald Rumsfeld has called these things the unknown unknowns - the things that we don't know while remaining unaware about the very fact of not knowing.
SQLite to store data. Web framework to interact with user. Python to glue them together. The need in all these parts comes naturally, you could not "forget" to use any one of them. Yet they are not the only parts that are needed.
Nowhere in Python's database API documentation it says that composing each query individually is highly inefficient in a database-driven application. No one hints that there exist a whole other class of libraries called Object-Relational Mappers (ORM). And I was not clever enough to deduce that on my own.
But I wasn't too dumb either. I figured that repeating myself every time I needed to run a simple query was wrong. So I stashed that code away into SQL class. I figured that Book and Author objects require a lot of common methods to interact with database. So I separated that code into TableEntityWithID class. I have essentially implemented an ORM without knowing what ORM is. Of course, it is far inferior to SQLAlchemy and the likes. Of course, I will never use it in another project now that I know that there are industry standard solutions in this area. But I see no point in replacing my (suboptimal) implementation now. Because it works and nothing is broken and nothing is missing.
I consider my ORM to be a valuable educational experience even if it somewhat stalled the overall progress of the project.
While I was developing the application I was using Bottle's builtin wsgiref web server. Running that server in production environment is not recommended because it was not built to withstand all the nastiness of open Internet. It is also a single threaded server which means it can not accept a request until it's done serving the previous one. The single-threadiness did not concern me - my app was intended to be used by two or three users tops and simultaneous requests would happen very seldom. Yet all the proper web servers (Apache, Nginx etc) are created to scale up to thousands and millions of users. Of course, they are multithreaded and multiprocessing. My application was not ready for that. Some would say SQLite is not ready for that either.
At first I considered crippling Apache to use only a single thread or switching to another server that can operate in a single thread mode. OpenShift seduced me with maintenance-free setup of Apache and I've abandoned that idea.
The code I've come up with to use a different instance of some objects for each thread is rather ugly. And it works only 90% of the time. So I just restart the Python workers every hour to avoid HTTP 500 errors when the database gets locked up. Not the best engineering decision of mine.
There probably exists a proper solution to multithreaded SQLite access, but so far I have not guessed what keywords to ask Google for it.
I was aware that changing database schema after the database was populated is hard and can lead to data loss. So I have put a lot of thought into designing it. I've even fired up GraphViz and created a nice chart that I've printed out and looked at on the commute. I wanted the database schema to be iron-clad and to require no changes in future.
I was very naive.
Of course it would require changes. Everything around us changes, we change, our expectations towards software change. At that point I understood that updating database manually was not a solution: there would be no trace left whether the schema was updated, which changes were introduced and which version does it correspond now to. So I've coded a simple transition module that would do that work for me. And later I've learned that the process I was automating is called database migration, and there are tools written by professionals for it.
One might think that splitting code into modules comes naturally. At least I did. And I was wrong. When I was not specifically thinking about their size some modules tended to grow large. A thousand lines is pretty hard to do a quick overview on, and even harder to split after the fact. Some modularity has to be designed from the beginning.
I don't know how I could've avoided that. Guess it comes with the experience.
This is another example of a sensible principle that's hard to come by without someone teaching you. Model-View-Controller is a logical extension of modularity principle which I was totally unaware about. I have made some intuitive steps in the right direction, but overall my application is a soup of interleaved components. That complicates maintenance and further development and makes the code more difficult to understand for other developers.
If I would've gone with a bigger framework like Flask or Django, MVC mindset might have been forced down on me. For a newbie who doesn't know anything some dictatorship isn't that bad.
After all the difficulties and complications (both expected and unexpected) I can proudly say that the Library project is a success. The website is up and running for almost a year. There has been no significant downtimes and no data loss, most of our books have been catalogued. And most important, me and my wife do enjoy using it!
The application supports:
- Creating and editing book entries.
- Storing and displaying book metadata, cover thumbnail and arbitrary related files. Each book can be connected with any number of authors, series and/or tags.
- Using ISBN to fetch book information from several third-party sources
- Queuing ISBNs for input. This is helpful when you process a lot of books with barcode scanner and don't have the time to clean up automatic metadata on each one of them
- Adding 1-to-5 star ratings and text reviews to any book in the library
- Searching for books by exact metadata match and with wildcards
You can access the source code on GitHub and see the site in action at https://morebooks.ml (registration is for family members only, sorry). Of course, there are plenty of improvements to be made (you can see how long the TODO list is), but the maintenance itself requires almost zero attention now and I can happily switch from being a developer to becoming the end user.