Realm is a mobile database: an alternative to SQLite and key-value stores
github.com
github.com
So what changed? Did the idea of a general purpose database system break because the usage scenarios became so diverse that we can no longer pretend that one size fits almost all? Is this just a perception thing because everybody started uploading his weekend hack toy database to GitHub and blogs about it?
And what really irritates me the most, database systems are really complex pieces of software requiring years and years of development time, where does this time come from and who pays for it? Especially having more and more different databases system to choose from does not really promise that developing yet another one will be an easy commercial success.
I for one think that it is awesome that it is now available for node.js as well. Especially with the new replication functionality that allows you to share live data between node instances running on different machines.
Those databases failed not necessarily because the technology was wrong, although they certainly made some choices that didn't help; those databases were often tightly coupled with opaque binary object serialization formats (with no way of looking into a database without going the binary route) that were in turn tightly coupled with the app's class hiearchy. For example, several of these actually post-processed your code (e.g. your Java bytecode) to insert serialization code, so that when you followed a relationship (e.g. "library.books[0].author"), it would automatically query under the hood and return the author as an object. They ended up being clunky to work with as a result; for example, in the beginning, very few of them had a query language, so the only way to query the data was by following relationships through your graph of objects.
Realm looks interesting as a new approach to the object database design, but at the same time seems a bit limited. It doesn't seem to have schemas or any kind of role-based permissions, and seems to make the same mistake of not including a query language (so you can't build a REPL for it, and every language binding needs its own query builder). The transaction semantics also seem to be entirely undefined (is it read-committed?). From what I can tell, it's also "offline first", in the sense that you have a local database that can sync to/from a server, but you're technically always working on an offline copy of the master data; while nice to have for a mobile app, it does limit what you can use it for. I also don't see any support for fine-grained updates. Every write you do seems to overwrite, which I assume means the last write always wins; how do you deal with conflicts, or implement things like counters, maps or arrays that need to be incrementally updated?
It also has a query language. But as one of the core design principles is to be as native to the language and platform as possible, the query language takes different forms. In Java it's a fluent API, in Cocoa it's NSPredicates string based. And the JS string based is roughly identical to that as well.
It's absolutely designed to be offline first. I'm curious as to which limitations you see with that?
Operational Transform is used to resolve conflicts at the property level, so I would say it's much more fine-grained than most other systems that would overwrite entire objects. The specific resolution rules are very simple and intuitive and how to deal with custom needs are described here: https://realm.io/docs/realm-object-server/#conflict-resoluti.... There is ordered lists already and explicit counters are being exposed in the SDKs at the moment.
You might find this article useful as well: https://realm.io/news/eventually-consistent-making-a-mobile-...
(By query language I don't mean an opaque builder based on NSPredicates or whatever, I mean a query language, written as text, which is portable across clients.)
OT is great, but it doesn't have that until it's actually exposed. My pet use case is string edits, which absolutely need to be cleanly conflict-resolvable.
The inconvenience with offline-first is that by definition you're never operating on current data: You have to sync to a central server before the data is visible to others, and you're editing potentially stale data. The latency of inbound syncing is particularly dependent on the volume of changes being sent from upstream, which might require that a client, so as not to be overwhelmed with updates, subscribe to a very small subset of the entire dataset. For example, if you open up a UI that edits a single document, you might want to subscribe to changes to only that document, in order to update the UI in real time. I don't know if that's even possible with Realm, or if you can in fact run it in "classic client/server" (i.e. "online first") mode.
* Interactive REPL (think psql)
* Use outside of C#, e.g. small bash scripts for automation of small tasks such as deleting old records in a cron job
* Copy/paste a into Slack, Github issues etc. without having to drag along any dependent code or worrying that it's not a complete query
Query builders are nice, but eventually they have to compile down to something "neutral" and portable.
I think most would agree there is no longer a one size fits all (or even most).
Realm (and I know someone will correct me if I'm wrong) started out as an on-device only replacement for relational-style dbs focused on ease-of-use and performance.
It's not the only one. Couchbase Mobile is a NoSQL stack that has an on-device db that's also easier to use, and has focused for many years on getting sync right (which is a hard problem).
It also times some time getting use to -- my biggest gripe is that the notification API won't tell you what actually has been modified, just that an object changed _somehow_. That makes a pure reactive design quite hard if those changes have non-trivial consequences, and not just a quick re-rendering of some element on the screen.
Looking forward to finding time to go back to SQLite where I can actually do complicated queries, relational joins, triggers, etc on a fast and mature code base.
But in general I rely on RxJava to provide fine grained notifications. For example (psudeocode):
someUserQueryThatReturnsObservable()
.map(user->user.firstName)
.distinct()
.subscribe(newFirstName->newFirstName)
It's messy because you end up having to repeat everything after map for every property, not really an ideal solutionI personally think they did an awesome job with the change api
Over the past year, it's become a lot more stable and I've seen far fewer user reported issues with it. As others have pointed out, it's also a very nice API and very simple to get set up.
The Realm team has also been very helpful in my experience and usually have responded to my support requests almost immediately, despite the fact I'm using their free product and not the DB-as-a-service they recently shipped.
Android support is coming in a matter of weeks, so stay tuned.
The docs you sent me seem to suggest that it's still just indices for objects modified. Where are the property changes?
> All changes to an object (addition, modification and deletion) must be done within a write transaction.
> Write transactions incur non-negligible overhead - you should architect your code to minimize the number of write transactions.
So is there a single write lock? If you try to start two transactions, will the second one be delayed or retried? Can reads proceed while the transaction is processing?
> Results instances are live, auto-updating views into the underlying data
So if I am doing x.first, x.second while the object is being updated in a second thread, will I get mixed state?
Is database crash proof? If the phone runs out of battery while in write transaction, will the database get damaged?
... so many questions
> So if I am doing x.first, x.second while the object is being updated in a second thread, will I get mixed state? So no you will not get mixed state.
> Is database crash proof? If the phone runs out of battery while in write transaction, will the database get damaged?
Yes it is crashproof. Fully ACID.
Any other questions?
This article goes a bit more into the details of threading: https://realm.io/news/threading-deep-dive/ And this one more into the core design: https://realm.io/news/jp-simard-realm-core-database-engine/
It's been really nice to use so far for me, plus the Realm team puts out a lot of helpful deeper dive blog posts: https://realm.io/news/
It seems there have been attempts at removing Android dependencies but it seems they were never finished, meaning you can't have it running on desktop apps.
This is (sadly) a showstopper for me.
Sometimes I just want a simple persistent on board storage mechanism for my mobile apps without all the problems that localStorage exhibits. I wonder if Realm could be a lightweight alternative?
The former is, as the HN post suggests, basically a database for mobile applications which you can use in place of SQLite or as a key value store.
The latter combines the database software which runs on the mobile devices with the Realm Object Server[2] which runs server side to provide something like Firebase (disclaimer: I've used the Mobile Database but haven't used the Mobile Platform (and Object Server) yet).
[0]: https://realm.io/products/realm-mobile-database/
https://realm.io/docs/javascript/latest and https://realm.io/docs/javascript/latest/api
Underneath it's LevelDB.