Hi there. Fortunately for the ML team, the Previews team kept a detailed cost breakdown for different preview types - including on the fly generation cost, async generation cost, and cost to serve a preview. Our ballpark accounted for some of the varying costs, though the difference is not significant.
Hi joosters, thanks for the question. It is always a good practice to ask if we are solving the right problem.
The decision to prewarm is ultimately a product decision to give users a better experience across the many surfaces where they encounter file previews. There are limitations to how fast previews can be generated on the fly, even under optimal conditions (unlimited top-of-the-line hardware, max of 1 request at a time). For instance, optimal on-the-fly preview generation for more expensive files (say, a video that is an hour long or a gigapixel image) can add 10s of seconds to tti -- not good from a user perspective!
Given this constraint, we wanted to optimize where we spend on the preview generation infrastructure without negatively impacting the user experience. We chose to do this with a combination of heuristics and the ML solution described in the article.
Couldn't you just load the first frame of the video, and aren't most image formats optimized for fast thumbnailing? I take your point that there are times when it will be slower, but are you saying that you assumed ML would be faster than on the fly, or did you actually check?
Hi mehrdada. In the article, we discuss how to evaluate tradeoffs of ML projects. One of these tradeoffs is cost of development and deployment vs. cost of not developing a solution. In our particular case, the tradeoff made sense.
Hi folks, author here. I am very excited to share this post about how we use machine learning at Dropbox to optimize our document preview generation process. Enjoy!
I'll be online for the next hour or so; happy to answer any questions about ML at Dropbox.
I work in an innovation ML-oriented lab, and we have a hard time identifying use cases with real added values.
So I wondered: who had the initiative to use ML in riviera, the riviera team or the ML team? How do you collabore between the two teams/worlds (production team and data science team)?
Hi setib, great question. The original idea to use heuristics for preview cost reduction came out of a Hack Week project. This led to an initial brainstorm meeting between the ML team and the Previews team about what this might look like as a full-fledged ML product.
From the beginning the ML team's focus was on providing measurable impact to our Previews stakeholders. One thing that helped us collaborate effectively was being transparent about the process and unknowns of ML (which are different from the constraints of non-ML software engineering). We openly shared our process and results, including experimental outcomes that did not work as well as planned and that we did not roll into production. We also worked closely with Previews to define rollout, monitoring, and maintenance processes that would reduce ops load on their side and provide clear escalation paths should something unexpected happen. Consistent and clear communication helps build trust.
On their side, the Previews team has been an amazing ML partner, and it was a joy to work with them.
I'm curious to know the answer to this question as well. I have done a fair bit of work working with organizations to identify ML use cases. When we looked at it from a business process perspective, honestly it didn't go very well. Trying to find company process specific interventions, especially in the format of building a funnel to priortize which to move forward, rarely surfaces unique or game changing ideas. We usually ended up generating a list of things where either ML played a minimal role, something more simple would have been better, or you'd need AGI.
What I've seen work better is a product approach, where ML is incorporated as a feature (rarely but possible the centerpiece) of a full solution for an industry that provides a new way of doing something and the value that comes with it. The caveat is that this is hard and takes up front R&D and product market fit research that any product would. It doesn't happen in a series of workshops with representatives from the business.
This Dropbox story is an obvious counterexample, and really looks like the mythical "low hanging fruit" that we always want to identify in ideation workshops. But I'd be careful trying to generalize a process for identifying ML use cases from it.
I’ve worked on ML across several large e-commerce firms, and two patterns I have seen along the lines of your comment:
1. many organizations dismiss ML solutions without actually trying them. Rather, if one hack week style prototype doesn’t work on the first try, it’s chalked up to “over hyped AI” and never sees the light of day. Organizations that succeed with ML don’t do it that way. Instead they ensure the success criteria are stated up front and measured throughout, so you can see why it didn’t work and iterate for v2 and v3. “We spent a bunch of R&D expense on magic bullet v0 and it didn’t succeed immediately” is a leadership and culture problem - you probably can’t succeed with ML until you fix that.
2. Many companies have no idea how to staff and support ML teams, and go through various cycles of either taking statistical researchers and bogging them down with devops or taking pure backend engineers and letting them do unprofessional hackery with no clarifying product quality ML expert in the loop.
You need a foundation of ML operations / infra support that cleanly separates the responsibilities of devops away from the responsibilities of model research, and you must invest in clear data platform tools that facilitate getting data to these teams.
If an org just figures they can throw an ML team sink or swim into an existing devops environment or they can require an ML team to sort out their own data access, it’s setting ML up for disaster - and again you’ll get a lot of cynics rushing to say it’s failing because ML is just hype, when actually it’s failing due to poor leadership, poor team structure and poor resourcing.
Yeah, for one thing, the scale of Dropbox may make this a uniquely worthwhile investment for them. Many apps have similar kinds of speculative caching features that could be optimized with predictive modeling, but the same cost-benefit analysis might show that it saves $1.7k/year instead of $1.7M, less than the cost of the feature's development and maintenance.
@andy99, @setib: we're a boutique that helps large organization in different sectors and industries with machine learning. Energy, banking, telcos, retail, transportation, etc. These organizations have different maturity levels and their functions expect different deliverables.
The organizations range on the maturity level from "We want to use AI, can you help us?" to "We have an internal machine learning and data science team that's overbooked, can you help?" to "We have an internal team, but you worked on [domain] for one of your project and we'd like your expertise".
For the expectations, you can deal with an extremely technical team that tells you: I want something that spits JSON. I'll send your service this payload and I expect this payload. So that's a tiny part.
Sometimes, you have to build everything: data acquisition, develop and train models, make a web application for their domain experts with all the bells and whistles, admin, roles, data management, etc. I wrote about some of the problems we hit here[0].
The point is that, finding these problems is an effort that requires a certain skill/process and goodwill from the clients. We worked on a variety of problems.
