"All the Monitoring" by Bryan Brandau from Best Buy: good overview of monitoring tools (other than Nagios). Showed Sensu, Graphite, Logstash. What the heck is up with all the different open-source dashboard projects, though? He mentioned at least five by name.
"Agile Financial Modelling" by the Investyr guys. A little weak, their main point was not to spend a bunch of time on useless stuff--- put a quick spreadsheet together spending 80% of your time on year 1, 15% on year 2, don't bother beyond year 3. I wish they had done more with showing what sort of scenarios you could play with once you had a spreadsheet in place, instead they actually built it on the fly.
"Do 'Data Science'" by a bunch of folks from NativeX. I walked out of this one--- should have gone to learn about Julia or attend the lightning talks instead, but both rooms were packed. The presentation was dull and process-oriented to start, and included an explanation of Bayes' Theorem that was tedious to those who already knew it and probably confusing to those who didn't. Really would have been a ton better if they'd started with "here's what we did for the company, want to learn how we did it" rather than some bogus "we follow the scientific method unlike you
Technical Support for Developers panel by Chris Warren, Luke Francl, and Kevin Whinnery. This was a lot of fun. The support they provide is mainly API support rather than purely end-user but they had good stories comparing different approaches (dedicated team vs. development team) and the value of direct customer feedback.
Shopping for Nerds by Neal Tovsen. This was a brief talk on how to weigh the risks of different technical "talent": outsourcing, freelancers, co-founder, etc. I think Neal accurately described the trade-offs, and his point that you need to make sure somebody is "looking out for you" is well taken. But the session could have been stronger with more examples to help illustrate what has worked and what hasn't--- we mainly got some (usually anonymized) horror stories.
I asked whether anybody in the room had actually been successful at throwing out their outsourced/cobbled-together MVP and building the "real" product afterwards. Nate Yourchuck said he's done a couple of consulting projects where he built a real product to replace emailed-around Excel spreadsheets. I remain dubious that you can really "build one to throw away" (pace Fred Brooks.)
I got to hear Neal give his pitch for "Apruve" several times today and yesterday, and I think there is quite a bit of potential there.
Journey to the Bottom of the Storage Stack by me. I thought it went pretty well, but I suspect I spoke too quickly. Also I kept wandering too close to another live microphone and causing feedback. My friend Scott brought a friend from Isilon along to keep me honest. That gentleman suggested an alternate proposal for an anomaly I'd found, of hot data blocks pointed to be cold metadata--- that frequently-written metadata blocks might never be flushed from the journal. I'm dubious, I don't know if ext4 actually implements that optimization.
Slides will go up on the Minnestar wiki within the next few days, hopefully.
Scaling with Cassandra by Jeff Bollinger + another NativeX (nee W3i) guy. Talked about why they chose to move from SQL Server to Cassandra, and about Cassandra's data and consistency model.
The numbers they gave for their production system: 12 nodes (each running 12 cores @ 2GHz) with 2x480GB commodity Intel flash drives. Per node they see about 240 writes/sec peak, and 888 reads/sec. (These numbers seemed low, I think they are observations rather than benchmarks.) They see about 3ms writes, 4ms reads (end-to-end) which is not too bad considering the extra network hop, but not blazing fast--- their SQL server achieves 1.5ms read latency, but most of the data is in memory after moving bulk data to Cassandra. Their solution is hybrid now, they kept SQL for the data that really was relational, and talked about the importance of understanding the data movement characteristics while doing the design. Jeff also mentioned that it took them a few times to get the hardware right, so they wished in retrospect they'd done more prototyping and experimentation using cloud resources instead of in-house boxes.
I always feel like I should stick around longer during the happy hour, but as usual I had had about all the monkeys I could take, and even on the edges of the large cafeteria area the conversational noise is quite high.