|
I have read the article and this thread with keen interest. I am curious to solicit community feedback on our choices as we faced the Node vs. Java choice in 2012. In particular, does our business ultimately suffer long term consequences by our stack choice? We think we ultimately have strong advantages vs. any competitor built using Node.JS, but the beauty of HN is that we can open up the thoughts for everyone to contribute. I am the CEO and founder of Codenvy. We make SAAS developer environments, and have been working as a team since 2009. Starting in 2012, we recognized the limitations of our initial system and needed to rebuild the entire system from the ground up. We had three criteria: performance, security, and modularity (of our core base and an ability for ISVs / enterprise to plug-in their own customizations). We raised $9M in Jan 2013, so we now have big investors and aspirations equally large. Our target market is primarily enterprises and ISVs that build Eclipse plug-ins. Individual developers and consultants are a secondary audience. We chose to go with a pure Java stack, after a careful evaluation of Node.JS vs. Java. We wanted to provide some of our thoughts, as we continually ask ourselves whether there are downsides to the approach we chose. For the record, we have competitors like Eclipse Orion and Cloud9 that are pure JavaScript / Node implementations, but we are - as far as I have seen - the only Java implementation in our market. Our team is currently 40 people, with 35 or so in R&D. Our engineers have experience with Java, Scala, JavaScript, AngularJS. Prior to this project, most engineers were working on server systems for content and document management firms. The hard stuff: sync, versioning, correlation, and so on. As for myself, my career started as a Java engineer in the 90s, but for the most part have been hands off as a technologist for a decade. My skills may have faded, but as a CEO, I am conversant. We did 2 prototypes on Node.JS where the client and server were combined systems. We ran a battery of performance tests and did conclude that we could probably run a large scale system on fewer physical nodes with Node.JS than with Java, but the gains were not enough to justify the choice. Our long term architecture was different from our competitors. Most SAAS dev environments provision each user / group their own VM, give you root access, and let you control that system. We chose an approach that would optimize overall system density by by offloading builds, debugs, and editing onto separate clusters. User environments will be virtualized and routed to the appropriate cluster on-demand. The user will feel like they have a personal VM, but our elvish helpers route traffic between the nodes across performance queues. After offloading LDAP (for identity), event / log management (for analytics), and cloud administration, the only core systems left was our API services (account / builder / runner / auth) and editing. These core systems would be subject to heavy iterative development, and Node's architecture could offer higher density of requests / users than Java, but by our tests the comparison was 10-20% difference, not 300%. By way of scaling, for every 10,000 concurrent developers, we average 20 editing nodes (I/O constrained), 75 debugging nodes (memory constraining), 5 builder nodes (CPU constrained). A similar TS / VDI implementation would require nearly 5,000 nodes of the same shape / size to service the same population. So, we saw a disadvantage to Java as we could have reduced our editing node even further. But we chose Java because:
a) The library selection. Broad and mature. In particular, it's data structures, file management, REST services.
b) GWT. This library gave us cross-browser portability of code written in Java.
c) Eclipse Plug-Ins. We knew we wanted to target Eclipse plug-in developers and saw the Orion project as optimized for Web dev, leaving a gap in Eclipse. We saw it as a business gamble that Eclipse devs who needed to move their plug-ins would find a pure Java plug-in architecture that was similar to RCP more attractive than a newer (but have to learn new tech & rewrite) approach.
d) Our history. We did a straw poll of our engineers before we started. We asked them if they felt that they would rather do Node or Java, and 75% said that they would enjoy Java more. When asked why? They had tool familiarity, were confident in the maturity of maven, and better job prospects. This last one completely surprised me.
e) JVM. The number of optimizations of the JVM would give us advantages for enterprise deployment. We built our system to be automatically installable on-premises. We do so with Puppet, but recognized that we would be encountering many IAAS and PAAS standards internally. Enterprises had security and performance tested the JVM, so we saw an opportunity to win hearts and minds of system admins, datacenter architects, and security specialists by our choice of architecture. So, we carried forward. Since we started working on our next generation architecture, we have focused hard on the best practices associated with agile, devops, and continuous delivery. We have 7 teams of 4-6 people, each focused on a different aspect of our stack. We are able to release any Jira issue at any point in the day. Each team can push into acceptance, staging or production without coordinating with the other teams. At the same time, we have achieved modularity and a rapid development tempo. While we haven't released our new system to the market yet, we find that the performance of the environment is similar to a desktop environment, and it scales cleanly with simulated load tests. We have also been able to give out our SDK (github.com/codenvy/sdk) to 6 different partners, some of which are Eclipse plug-in developers, who are able to code, build, test, integrate, and publish their plug-ins without complicated dependencies to our core team. Their feedback was that the port of their plug-in took about 1/2 the time of the original development. And there is no real way to estimate what the effort would be to do the same in a Node-based offering like Orion. What we can't measure is whether our competitors are able to release faster & more functionality than we are. It seems like a crap shoot. One other variable we can't account for is the cost to maintain legacy. The system we have Codenvy.com is legacy for us, and will be retired in Q3. We have to continually allocate 3-5 people to work on bugs, special partner projects, or other related items to keep that system live while we also work to roll out the new system. We know this is a drag, and some of our engineers think we may more than double our productivity once we consolidate to a single architecture. I am a bit skeptical at that - but I find that it's unwise to disagree with my engineers. I tend to lose. So at this point, we feel like we made the right choice. But time will tell - we begin selling our enterprise offering earnestly once we ship our new system in Q2. Looking forward to the discussion. |