Sunday, October 29, 2017

HPTS'17 day 1

This post is a continuation from yesterday's post. I cover only some of the talks, you may check the other talks from the HPTS agenda.

Mind your state for your state of mind (Pat Helland -Salesforce)

HPTS day 1 started with a keynote from Pat Helland. Pat Helland is a database veteran. I had covered some of his papers on this blog before. He writes insightful position papers with a unique style. The keynote provided an overview of trends in storage and computing and hit the high notes from his earlier position papers as well, and mentioned this table:
                          | fast-reads | fast-writes | read-your-write
linearizable       | no             | no              | yes           
nonlinearizable | yes            | yes            | no             
cached-data       | yes            | no             | no             

The point is you can't have everything. Pat argues "immutability is a solid rock to stand on" and is the closest you can get to yes, yes, yes, in that table. However, Pat cautions that what matters is not hitting yes, yes, yes on that table. Don't fall in love with any of those consistency semantics, what matters is the application requirements. Different applications demand different behaviors from durable state: Do you want it right or right now?

Here is Pat's presentation slides.

Stop building infrastructure (Evan Jones, Bluecore)

Evan started the presentation with a disclosure: warning: zero facts ahead. This was sign that this will be a deliberately provocative/controversial position. He argued that developers should stop building infrastructure (i.e., databases, load balancers, service orchestration software).

His main point was that companies already have their workflows/toolchains streamlined on a cloud vendor and compared to the simplicity of integrating an existing service from the cloud vendor to their workflows/toolchains, it is an uphill-battle to overcome when integrating a 3rd party to their workflow. Ain't nobody got time for that.

His other points were:

  • the cloud vendors are better at building infrastructure than you
  • even if you succeed, your infrastructure is going to run at cloud, and will make money to the cloud vendors
  • cloud vendors can destroy you via copy and crush
  • at a company it is easy to justify paying an existing cloud vendor, but it is cumbersome to convince management to buy 3rd party software

These are fair points, but I think his talk should instead be making the opposite point: things are dire for the infrastructure startups when facing the cloud vendors, so how can we help/encourage them? At one point in his talk, he mentioned that their company used only one outside vendor, the datadog, because google cloud hiked up the prices on its monitoring service. This shows that we should be rooting for the success of infrastructure services/companies so that the cloud vendors don't grab all power against the developers.

Moreover Evan's points don't take into account enterprise markets and also foreign markets like Europe and China. Infrastructure companies do exist because there is market for them. Several infrastructure/databases companies have done really great financially, and I hope we will see more success stories from them in the future as well.

Blockchain and Databases - C Mohan (IBM)

Mohan gave an overview of blockchain and talked about the HyperLedger project. His talk(s) and more information can be found at http://bit.ly/CMbcDB.

CockroachDB: From OLTP to HTAP - Arjun Narayan (Cockroach Labs)

Here are the slides for the talk. OLTP stands for online transaction processing, and HTAP stands for hybrid transactional/analytical processing.

The talk gave a quick overview of CockroachDB's scale-out architecture to implement a WAN distributed SQL database service. When building CockroachDB the team prioritized correctness and stability over performance, and attacked those first. They are now working on performance improvements and achieving scale-out OLTP. Arjun talked about how HTAP can be done over CockroachDB in the future. He likes to leverage on the CockroachDB OLTP layer and adopt incrementally updated materialized views as the building primitive. Concretely, he argues that since CockroachDB timestamps give serializable transactions that span OLTP+OLAP, it is possible to build dataflow systems like Naiad or timely dataflow for achieving HTAP.

Gongshow

After the dinner, we gathered at the chapel for a potpourri of 5 minute talks. Andy Pavlo's student Joy Arulraj organized the session. (Joy will be on the academic job market and he will be a great catch for whoever manages to hire him.)

There were many interesting talks in the gong session. And with 5 minutes per talk, they were fast-paced as well.

Joy also convinced me to present at the gong show at around noon, so I had to write some slides in the afternoon. Thanks to the emacs-org-mode export to beamer I was able to come up with slides in time for the talk. I presented about our ongoing work on wPaxos, our WAN Paxos protocol. I will be writing a blog post about that sometime soon.

No comments: