SIGMOD DEEM 2018 Panel Discussion

The ACM SIGMOD Second Workshop on Data Management for End-to-End Machine Learning (DEEM) was successfully held a few weeks ago in Houston, TX. The goal of DEEM is to bring together researchers and practitioners at the intersection of applied machine learning (ML) and data management/systems research to discuss data management/systems issues in ML systems and applications. This blog post gives an overview of DEEM'18 and a lighthearted summary of the exciting and informative panel discussion.

Overview of DEEM 2018

As per the SIGMOD Workshops chairs, DEEM'18 had 117 registrations--almost half more than the next largest workshop at SIGMOD'18, about thrice as large as a typical SIGMOD/VLDB workshop, and likely the highest in the history of SIGMOD/VLDB workshops! Clearly, this is a red-hot area and our program had stirred the curiosity of a great many people. The day of, "only" about 70-80 people showed up. Thanks to sponsorship from Amazon and Google, we also funded 4 student travel awards. There were 10 accepted papers, with 3 presented as long talks and 7, as short talks. These papers spanned many interesting topics, including new ML programming models, scalable ML on DB/dataflow systems, human-in-the-loop ML exploration tools, data labeling tools, and more. A variety of top schools and some companies were represented. All of this was made possible thanks to the hard work of a top-notch PC. We had 4 excellent invited keynotes/talks from both academia (Jens Dittrich and Joaquin Vanschoren) and industry (Martin Zinkevich from Google and Matei Zaharia with his Databricks hat).

Jens shared his thoughts on what DB folks can bring to ML systems research and education, as well as a recent line of work on using ML to improve RDBMS components. Joaquin spoke about his work on ML reproducibility, collaboration, and provenance management with the successful OpenML effort. Martin gave a unique talk on a topic seldom addressed in any research conference--how to navigate the space of objectives for ML-powered applications before even getting to the data or ML models. He used a very Google-y example of improving user engagement via click measurements. Finally, Matei spoke about the recently announced MLFlow project from Databricks for managing the lifecycle and provenance of ML models and pipelines.

Panel Discussion

Getting to the panel discussion itself, the topic was a hot-button issue: "ML/AI Systems and Applications: Is the SIGMOD/VLDB Community Losing Relevance?" I moderated it and my co-chair, Sebastian Schelter, also helped put together the agenda. To make the discussion entertaining, I played the devil's advocate and made the questions quite provocative. Apart from the 4 invited speakers, we had 2 additional panelists: Joey Gonzalez (faculty at UC Berkeley) and Manasi Vartak (PhD student at MIT), all of of whom are working on DEEM-style research. Two previously confirmed invitees, Luna Dong and Ce Zhang, were unfortunately unable to make it to the workshop.

Photos from the workshop. L to R: (1) The DEEM audience. (2) The Panelists: Matei, Joaquin, Jens, Joey, and Manasi (Martin not pictured). (3) Advocatus Diaboli.


First off, I clarified that the "irrelevance" in our question was meant only in the context of ML systems/applications, not data management in general, eliciting laughter from the audience. After all, as long as there is data to manage, data management research is relevant, right? :) But with the dizzying hype around AI and deep learning, we saw the above question as timely. The discussion was supposed to have 9 questions across 3 topics--problem selection/research content, logistics/optics of publication venues, and student training. But the amount of discussion generated meant we could cover only 6 questions. I started with an overview of the history of ML systems, from SAS and R in the 1970s all the way to the DEEM community today. In the rest of this post, I summarize each question, its context/background, and the panel responses and discussion. For brevity sake, I will not always identify who said what.

We started with two fun rapid-fire questions that I often use to put my students on the spot and gauge their technical worldview. What is a "database"? What is a "query"? The Merriam-Webster Dictionary says a database is just an "organized collection" of data, while a query is just a "request for information" against a database. Interestingly, almost all the panelists said similar things although the (wrong) definition that "a database is a system for managing data" did come up once. Most considered a query as a "program" run against a database. No relational. No structured. No system. Not even logic. The panel was off to a flying start!

Q1. Is "in-database ML" dead? Is "ML on dataflow systems" dead? Is the future of ML systems a fragmented mess of "domain-specific" tools for disparate ML tasks?

Context/Background:
There has been almost 2 decades of work on incorporating ML algorithms into RDBMSs and providing new APIs to support ML along with SQL querying. This avoids the need to copy data and offers other benefits of RDBMSs such as data parallelism. Alas, such in-RDBMS ML support largely failed commercially, according to Surajit Chaudhuri, a pioneer of in-RDBMS ML tools. At XLDB'18, he explained that a major reason was that ML users wanted a lot more tool support (latest models, iterative model selection, complex feature processing) that SAS and similar products offered. Those vendors also recognized the importance of near-data execution and connected their tools to RDBMS servers, leveraging user-defined functionality of such systems to reduce data copying. Moreover, statisticians and data scientists were unfamiliar with SQL and preferred the familiarity of SAS, R, and similar tools, sealing the fate of in-RDBMS ML support at that time. That said, as storage became cheaper, many enterprises no longer mind copying data to Hadoop clusters and using Mahout for ML. Anecdotally, some users also do this to reduce the load on their costly RDBMSs, which they use mainly for OLTP. Spark MLlib and similar "ML on dataflow systems" are now largely replacing Mahout. But in this "era of deep learning," the programming and execution architectures of both RDBMSs and Spark-style systems seem highly inadequate. So, most users stick to in-memory Python/R for standard ML and TensorFlow/PyTorch for deep learning. Thus, tackling data issues in ML workloads typically requires problem-specific tools. Perhaps ML is just too heterogeneous for a unified system. Hennessy and Patterson recently said that the future of the computer architecture community is in disparate "domain-specific architectures." Should the DEEM community be content with a similar future for ML systems with no unifying intellectual core like relational algebra/SQL for RDBMSs?

Panel Discussion Summary:
By design, this question induced a sharp polarization among the panel and set the tone for the rest of the discussion. Martin and Manasi agreed that in-database ML is pretty much dead and that custom problem-specific ML+data systems is the inevitable future. Manasi also opined that there will never be an equivalent of SQL for ML in the foreseeable future. Jens countered that even if RDBMSs are dead as an execution engine for ML, relational-style ideas will still be relevant for new custom ML systems, which everyone agreed with. Matei opined that while in-database ML is dying, if not already dead, ML-on-dataflow-systems is alive and well, since he finds many enterprise customers of Databricks adopting Spark MLlib. Joey weighed in based on his experience spanning both ML-on-dataflow and custom ML systems that both kinds of systems will co-exist, albeit with more emphasis on the latter. The overall consensus was that the application space for ML systems is indeed quite fragmented with different operating constraints and environments dictating which tools people will use. Most panelists agreed that while in-database ML may no longer be a particularly promising research direction, ML on dataflow systems, which is a more general environment than RDBMSs, are still a promising avenue for new ideas and will still matter for certain kinds of ML workloads.

Q2. What are the major open research questions for the DEEM community? Is the DB community's success with RDBMSs a guiding light or just historical baggage?

Context/Background:
Many CS communities are undergoing "crises" and witnessing massive "paradigm shifts," to borrow Thomas Kuhn's famous words. Perhaps the best example is the natural language processing (NLP) community. Almost 3 decades of work on feature engineering for applied ML over text have been discarded in favor of new end-to-end deep learning approaches. Some NLP folks say they find the new paradigm refreshing and more productive, while others say they had to take therapy to soothe the trauma caused by this upheaval. In this backdrop, the DEEM community can approach research on ML systems using RDBMSs as a "guiding light" to tackle problems and propose ideas. But this philosophy is fraught with the infamous pitfall of the streetlight effect--only problems/ideas that are easy to connect with RDBMS-style work and/or appease "the RDBMS orthodoxy" will get attention instead of what is truly valuable for ML applications. This pitfall is a highway to practical irrelevance, wherein researchers publish papers that look good to each other, while the "real world" moves on. An alternative philosophy is a clean slate world view in exploring novel problems/ideas in ML systems. But this philosophy is fraught with the risk of repeating history, including wasteful past mistakes in data systems research. Is it even possible to get a judicious mix of both these philosophies?

Panel Discussion Summary:
The first part of the question elicited a wide range of responses. Overall, several major open research problems/topics were identified by the panel, especially the following:

• More support and automation for data preparation and data cleaning pipelines in ML
• Abstractions and systems for ML lifecycle and experimentation management
• Efficient ML model serving and better integration of ML models with online applications
• Better visualization support for debugging ML models and data
• Frameworks to think about how to craft ML prediction objectives, especially beyond supervised ML

The second part of the question was met mostly with a measured response that RDBMS ideas will still matter in the context of ML systems but we need to pick and choose depending on the problem at hand. For instance, the DB community has long worked on data preparation, ETL, and data cleaning. But adapting them to ML workloads introduces new twists and requires new research in the ML context, not just routine application or extension of DB work. Joey and Martin also cautioned that it is important to study ML systems problems in the context that matters for ML users and developers, which might often require departing entirely from RDBMS-style ideas. The operating/distributed systems community routinely witnesses such changes. But it is likely that such changes will cause painful "culture shocks" for the DB and DEEM communities, given the stranglehold of the successful legacy of RDBMSs.

Q3. Why has 30yrs of work in the DB community on ETL and data cleaning had almost no impact on data preparation for ML among practitioners?

Context/Background:
Recent surveys of real-world data scientists, e.g., this massive Kaggle survey and this CrowdFlower report, repeatedly show that collecting, integrating, transforming, cleaning, labeling, and generally organizing training data (often collectively called "data preparation") dominates their time and effort, up to even 80%. Clearly, this includes data cleaning and integration concerns, which the DB community has long worked on. And yet, none of the data scientists interviewed or anecdotally quizzed seem to have found any techniques or tools from this literature usable/useful for their work. To paraphrase Ihab Ilyas, a leading expert on data cleaning research, "decades of research, tons of papers, but very little success in practical adoption" characterizes this state of affairs. Perhaps data cleaning is just too heterogeneous and too dataset-specific--more like "death by a thousand cuts" rather than a "terrible swift sword." Perhaps there are just way too many interconnected concerns for a generic unified system to tackle, whether or not it applies ML algorithms internally. What hope is there for DB-style data cleaning work in the ML context when its success in the SQL context is itself so questionable?

Panel Discussion and Summary:
Naturally, the provocative phrasing of the question elicited smiles and raised eyebrows, as well as a heated discussion. I separate the discussion and summary to highlight the different perspectives.

Jens countered that the DB community's work on data cleaning, especially those that apply ML, have indeed had an impact or at least, look very promising. Joey opined that part of the problem is with the term "cleaning." A lot of time is inevitably spent by practitioners on understanding and reshaping their data to suit their tools, to bring in domain knowledge, etc., but all such activities get grouped under the catch-all term "cleaning." Manasi agreed, adding that recasting the data representation for the ML task in peculiar ways also gets talked about as data cleaning or organization. Moreover, almost no ML or data mining curricula teach such data cleaning/prep issues, which skews perceptions. Jens agreed, adding that the "data cleaning" area has a major marketing problem, since it sounds so boring and janitorial. He suggested a clever play on words for naming: "bug data analytics"!

Joaquin said that it will be hard to eliminate humans completely and better human-in-the-loop solutions are needed to reduce manual effort. Jens suggested that pushing more of the cleaning steps into the ML modeling itself could help, similar to what some tree-based models already do. Joaquin agreed that making ML models more robust to data issues is also promising. Joey cautioned that ML will not be a panacea, since it relies on useful signals and enough of it being present in the data to achieve anything useful. One will still need domain expertise to guide the process and set the right objectives. Matei then interjected to opine that the DB community's work on SQL and dataflow tools has had a major impact on data prep and without this work, the 60% in the survey may very well have been 95% or more! Martin then pondered if an excessive focus on "cleaning" the data is wrongheaded when one should actually be "fixing" the data generating process, especially in Google-like settings, where most of the data is produced by software.

In summary, a major takeaway was that there is likely an unhelpful terminology confusion between researchers and practitioners in the data prep for ML arena, which could hinder progress. Another was to collaborate more with the ML community to make ML models more natively robust to dirty data. But there was consensus that data prep for ML, including cleaning/organizing/transforming data, will remain a core focus for the DEEM community.

Q4+Q5. Is the DB community too obsessed with (semi-)structured data and ignoring the deep learning revolution for unstructured data? On the other hand, is deep learning too overrated for ML analytics outside of "Big Tech" (Google, Amazon, etc.), especially for enterprises?

Context/Background:
These 2 questions are based on 2 key findings from the massive Kaggle survey, shown in the screenshots below.

"Relational" data includes both tables and multi-variate time series. Text data is clearly ubiquitous too, while images are not far behind. Given this, should the DB community be thinking more holistically about data rather than shoehorning themselves to structured and semi-structued data? Since deep learning with CNNs and RNNs is the way to go for text and multimedia analytics, the DEEM community should look at deep learning more. But the survey also shows that the most popular models are still linear models, trees, SVMs, Bayes Nets, and ensembles, way above CNNs and RNNs ("neural networks" in this list are likely just classical MLPs). This is likely related to the previous finding--relational data dominates their use cases and interpretability/explainability/actionability are crucial, not just accuracy. GANs, the new darling of the ML world, are at a mere 3%. Overall, there is a huge mismatch in what ML researchers consider "sexy" and what is important for ML practitioners! One might think deep nets are still too new, but the deep learning hype has been around for half a decade before this survey was done. So, one can only conclude that deep learning is overrated for most enterprise ML uses cases. This is a massive rift between the enterprise world and Big Tech/Web companies like Google, Facebook, and Amazon, who are using and aggressively promoting deep learning.

Panel Discussion and Summary:
Once again, due to the amount of interesting discussion these questions generated, I separate the discussion and summary to highlight different perspectives.

Manasi started off by agreeing that deep learning for text, speech, images, and video is something the DEEM community should study more but opined that current deep learning methods do not work on relational data. Thus, there is lots of room for work in this context. Joey bluntly stated that deep learning is indeed overrated and that most real-world ML users will remain happy with linear models, trees, and ensembles (e.g., RandomForest and XGBoost)! He also joked that we should rename logistic regression as a "deep net of depth 1." The compute cost and labeled data needs of deep nets are impractical for many ML users. Deep nets are also expensive to serve/deploy en masse. While deep nets are useful for speech and images, most users will just download and reuse pre-trained deep nets from Google/Facebook/etc. for such data, not train their own. Overall, his position was that surmising that the whole world will switch everything to deep learning is completely mistaken. For good jovial measure, he then added that his students are all working on deep learning for their papers!

Martin agreed that ML users should start simple but then also try complex models, including deep nets, based on available resources. For images and text, deep nets that exploit their structure are becoming unbeatable. Another key benefit of deep nets is that they are "compact" artifacts. Thus, their serving-time memory access characteristics are better than many other models that need extensive and cumbersome data transformation pipelines for feature engineering. These older models are also a nightmare to port from training to serving. Joey then interjected saying that such serving benefits only hold for "medium-sized" nets, not the 100s of layers ML people go crazy over. Matei weighed in with an anecdote about a Databricks customer. Deep nets are increasingly being adopted for images (and for text to a lesser extent), which are often present along with structured data. One example is "Hotels.com" using CNNs for semantic deduplication of images for displaying on their webpage. But he also agreed that not everyone will train deep nets from scratch. For instance, pre-trained CNNs can be used an image "featurizer" for transfer learning in many cases to greatly reduce both compute and data costs.

In summary, there was a consensus that the DB community should welcome more work on unstructured data and deep learning-based analytics. But there was also caution against getting carried away by the deep learning hype and levity that one should still work on deep learning anyway. The heterogeneity of ML use cases and requirements means that a diverse set of ML models will likely remain popular in practice for the foreseeable future.

Q6. Should the DEEM community break up with SIGMOD and join SysML? What more can we do to enhance the impact of DEEM-style work?

Context/Background:
The DB community's PC processes, as well as the kinds of research it values are now hotly debated topics. Stonebraker declared recently that SIGMOD/VLDB PCs are unfair to systems-oriented research. While praiseworthy steps are being taken by SIGMOD/VLDB to reduce such DB research culture wars, will this situation lead to DEEM-style work falling through the cracks? SIGMOD/VLDB PC chairs also repeatedly face the issue of not enough ML systems-related expertise/knowledge being available on their PCs and conflation of DEEM-style work with data mining work, worsening the reviewing issues. The SysML Conference was created recently as a home for ML systems research, including DEEM-style work, to avoid such issues. So, in terms of the potential for visibility, research impact, and fairness of research evaluation, is it better for DEEM to break up with SIGMOD and join SysML? What other venues are suitable for DEEM-style work?

Panel Discussion Summary:
This question drew both audible gasps and sniggers from the audience. It was also highly pertinent for at least 2 panelists. Joey helped run the MLSys/AISys workshop at NIPS/ICML and SOSP for the last few years and NIPS BigLearn for a few years before that. Matei helped start SysML in 2018 and is its PC chair for 2019.

Matei readily admitted that as a head of SysML, he will be delighted to have DEEM at SysML. According to him, each paper at SysML will get at least 1 expert reviewer each from ML and systems. The PC meeting will be in person. He opined that SysML is a good fit for DEEM, since DB-inspired work is a core focus of SysML. DB-style ideas can be impactful in the ML systems context as they were with "Big Data" systems. Moreover, since many well-known ML experts are involved with SysML, it offers more visibility among the ML community too. He also suggested that the smaller and more focused communities of NSDI and OSDI are other good options for DEEM-style work that is systems-oriented. SysML will be modeled on their processes. Joey countered that it is perhaps better to keep DEEM at SIGMOD to ensure ML-oriented work gets more visibility/attention in the DB world, which has a lot to offer. He also suggested the MLSys formula of colocating DEEM with 2 different venues, say, in alternate years. One could be SIGMOD and the other could be SysML or an ML venue like NIPS. He wondered if MLSys had overemphasized operating/distributed systems aspects of ML at the cost of other systems-oriented concerns. DEEM at SIGMOD can be a forum for ideas from all DB cultures and can be complementary to SysML.

Martin asked if it was harder to get DB/systems folks to work on ML concerns or vice versa. The latter is widely considered harder. Joey interjected to suggest that building strong artifacts aimed at ML users can improve visibility in the ML world, a la TensorFlow/PyTorch. Martin and Joaquin emphasized the need for solid standardized/benchmark datasets for DEEM-style work on data prep/cleaning/organization/etc. to be taken more seriously in the ML world. This is similar to the UCI repo and ImageNet, both of which boosted ML research. Manasi opined that DEEM-style work need not be so ML-oriented to be publishable at NIPS. But she suggested that SIGMOD/VLDB should create an "ML systems" track/area and add more ML expertise to their PCs. But focusing only on systems-oriented stuff in ML could lead to wasteful repetitions of other DB-style ideas. She said keeping DEEM at SIGMOD will help the DB world stay engaged. Matei had a caveat that without proper ML expertise on PCs, there is a danger of publishing papers that "look nice" but lack ML methodological rigor. While such papers will be discredited in the longer run, they will waste time/resources. Researchers interested in ML systems work should first understand ML well enough. The same holds for PCs. Martin concluded by saying that since a lot of ML systems research studies practical and industrially relevant problems, regardless of the venue, researchers interested in impact should talk to practitioners in industry.

In summary, while the panel was divided on whether DEEM should swap SIGMOD for SysML, they had good suggestions on increasing visibility and impact of DEEM-style work, including building good artifacts, dataset standardization, and bringing more ML experts to SIGMOD/VLDB/DEEM PCs. I closed the discussion declaring that as the DEEM organizers, we have no plans of leaving SIGMOD, since we are thoroughbred "database people," eliciting laughter from the panelists and the audience, likely in relief. :)

Concluding Remarks

Overall, the panel discussion was provocative and passionate but also insightful and constructive. Many of the audience members and panelists later opined that they too found the discussion educational and entertaining. The questions we could not cover included some industry trends and student training issues--another time then. A big thank you again to my fellow DEEM'18 organizers (Sebastian Schelter and Stephan Seufert), the steering committee, the PC, our invited speakers and panelists, the sponsors (Amazon and Google), the officials/volunteers of SIGMOD'18, and all the authors and attendees who made DEEM'18 such a success! I believe DEEM will be back at SIGMOD 2019.

Disclaimer:

We do not own the copyrights for the illustrations used in this article for educational purposes. We acknowledge the sources of the illustrations in order here: https://edtosavetheworld.com/2014/05/28/1-thomas-kuhn-the-structure-of-scientific-revolutions/, http://first-the-trousers.com/hello-world/, http://visit.crowdflower.com/rs/416-ZBE-142/images/CrowdFlower_DataScienceReport_2016.pdf, https://en.wikipedia.org/wiki/Lingchi#/media/File:Martyrdom_of_Joseph_Marchand.jpg, and https://www.kaggle.com/surveys/2017/. If a copyright owner wants their illustration removed, we are happy to oblige.

The Culture Wars of Data Management

Scientists are people. Despite their fervent protestations of objectivity, all scientists are prone to conflating subjective experience with objective truth, at least once in a while. Einstein himself infamously dissed on quantum mechanics initially but then contributed to its growth. The "whole truth" is an elusive and enigmatic beast. Consequently, people (including scientists) often fight "tribal culture wars" under the illusion that they (and those who agree with them) are right and all others are wrong. Perhaps nothing captures this issue more eloquently than the timeless parable of the blind men and the elephant. This issue arises over and over again in all fields of human endeavor, including the sciences. In this post, I examine how this issue is affecting the "data management" research community. My motivation is to offer a contemplative analysis of the state of affairs, especially for the benefit of students and new researchers, not prescribe definitive solutions.

This post is partly inspired by Pedro Domingo's article, "The Five Tribes of Machine Learning." He explains why there are deep "tribal" divisions within the machine learning (ML) community in terms of both the topics they pursue and their research culture. For instance, explosive battles often take place between the "Bayesians" (think statistics) and the "Connectionists" (think deep learning), two of ML's oldest tribes.

The data management field too has similar divisions, albeit not as strongly partitioned by tribes. It all came to a head last year when some prominent members of this community issued an ultimatum to SIGMOD and VLDB, the top data management research conferences. Criticizing the repeated unfair treatment of systems-oriented papers by reviewers (e.g., wrongly deeming a lack of theoretical analysis as low "technical novelty" or "technical depth," or crass dismissive statements such as "just engineering"), they threatened to fork off and start a new systems-oriented data management conference. Such brinkmanship is neither new nor unique to this field. In fact, break ups often occur on account of "unfairness," e.g., IMC splitting from SIGCOMM, ICLR vs CVPR, Mobisys vs MobiCom, the list is long indeed! From my own experience, I agree with the core claim of the ultimatum. Apparently, the SIGMOD/VLDB leadership also agreed. The PC chairs of VLDB then emailed the whole PC (which I was on) with guidelines on how to fairly evaluate systems papers. Problem solved? Haha, no.

Why write this post? Who is it for?

It may appear puerile to put labels on entire "research cultures." But such cultural gaps are ubiquitous, even in physics. Acknowledging and understanding such gaps is the first step to either mitigating them or making peace with them. Without such understanding, I reckon the data management community might be on a death spiral to more fragmentation (or maybe that is inevitable due to other factors such as size, who knows?). I see such labels as a way of acknowledging the different cultures, tolerating them, and even celebrating the intellectual diversity. More importantly, it is crucial for students and new researchers to be aware of such gaps and why they exist. This could help them make more sense of paper rejections or even negative in-person interactions with other researchers. From my experience, many established researchers transcend the cultural gaps and often bridge multiple cultures. Even among those that stay within one culture, many are not antagonistic to the others. I just wish everyone would be more like these folks. I hope this post nudges more people, especially newcomers, towards that ideal world.

The Four Canonical Cultures (as I see it)

Based on my interactions with dozens of researchers, reading hundreds of papers, and reviewing for SIGMOD/VLDB for the last two years, I see at least 4 "canonical" cultures (to overload a classic DB term!). Unlike Pedro's tribes of ML, which are partitioned by areas (e.g., Connectionists study neural networks), it is misleading to delineate the divisions within the data management community by areas/topics because many areas, e.g., query optimization, have all four cultures (or at least more than one) represented. Thus, my split is based on the inherent expectations of each research culture, their methodologies, the non-data management fields they get inspiration from (including technical vocabulary and "hammers"), and the style and content of their papers. Such differences are more insidious than areas, which is perhaps why the "culture wars" of data management are more damaging than the tribals wars of ML. These cultures are not mutually exclusive; in fact, many researchers have successfully hybridized these cultures and all 2^4 possible combinations are represented at SIGMOD/VLDB to varying extents. Anyway, here are the 4 canonical cultures I see:

• Formalist
• Systemist
• Heuristicist
• Humanist

The Formalist Culture

The Formalists bring a theory-oriented view to data management. Their culture has been influential since the founding years of the data management field, going back to Ted Codd himself. Formalists draw inspiration from theoretical fields such as logic, math, statistics, and theoretical CS, especially combinatorial optimization, formal methods/PL theory, and ML theory. They seek a high degree of formal rigor in the characterization and communication of ideas, especially in papers. Common elements in Formalist papers are non-trivial theorems and proofs, typically of the "hardness" of problems (couched in the language of computational complexity theory), approximation or randomized algorithms, and formal analyses of the complexity and quality of the algorithms. Many pride themselves in the theoretical sophistication of their results.

Some Formalists also publish in non-data management venues such as SODA, LICS, and NIPS, but curiously, not so much in STOC/FOCS. Database theory (think PODS/ICDT) is a major part of this culture, but not all Formalists are database theoreticians, i.e., some publish regularly at SIGMOD/VLDB too. Some Formalists have a reputation for expecting a high degree of rigor in all data management papers.

The Systemist Culture

The Systemists bring a systems-oriented view to data management. Their culture has also been influential since the founding years. Many pioneers of this field, including Charlie Bachman, Michael Stonebraker, Jim Gray, and David DeWitt can be considered as Systemists. They draw inspiration from the computer systems fields broadly writ (operating/distributed systems, compilers and software engineering, networking, computer architecture, etc.). Interestingly, their ideas have reshaped some of those other fields, e.g., the concept of ACID and distributed systems. Most Systemists seek a high degree of real-world practicality in ideas and care much less (compared to Formalists) for rigor in data management papers. Some even dislike seeing theorems. Common elements in Systemist papers are system architecture diagrams, discussions of system design decisions, and analyses of system trade-offs, typically with extensive experiments. They typically use system runtime or throughput metrics on synthetic workloads and datasets (e.g., TPC benchmarks). Scalability is often a key concern. Many pride themselves in the practical adoption of their systems, including via startups they found.

Some Systemists and Formalists have a reputation for waging "wars" against each other's culture or even other fields. The war between Bachman and Codd on the (de)merits of the relational data model is one example, as is the war between Stonebraker and Ullman on whether database theory research is needed at all. More recently, some Systemists have waged wars against the distributed systems field, e.g., this now-infamous blog post. Curiously, most Systemists tend not to publish in non-data management computer systems venues such as NSDI/OSDI/SOSP/ASPLOS/ISCA/etc.

The Heuristicist Culture

The Heuristicists draw inspiration from the fields of artificial intelligence (ML, natural language processing, etc.), but also logic, math, statistics, and theoretical CS. This culture started growing mostly from the 1990s. Popular topics where this culture is well-represented are data mining, Web/text data analysis, data cleaning, and data integration. Heuristicists design practical heuristic algorithms, typically without a high degree of rigor in the exploration of ideas (compared to Formalists) and without deep systems-oriented trade-off analyses (compared to Systemists). But there is substantial diversity in this culture. Many papers use rigor in communication and thus, seem closer to the Formalists. Others focus on complex algorithmic architectures and thus, seem closer to the Systemists. Many researchers bridge this culture with the Formalists, especially on the topic of data integration. But many problems in such topics are so "messy" that theoretical work alone goes only so far. So, many researchers also bridge this culture with the Systemists. Common elements in Heuristicist papers are math notation but no non-trivial theorems, large algorithm boxes or diagrams, and extensive experiments, typically with real-world datasets and workloads. They typically focus on quality metrics such as accuracy, precision, recall, AuROC, etc. Runtime or scalability metrics are less common.

Data cleaning and data integration have become central themes in data management research. But a large chunk of data mining researchers broke up with the SIGMOD/VLDB community and joined the newly created SIGKDD community along with applied AI/ML researchers. However, many data mining researchers still publish routinely at SIGMOD/VLDB along with SIGKDD/ICDM/etc. Similarly, many Web/text data researchers also publish at both SIGMOD/VLDB and WWW/AAAI/etc.

The Humanist Culture

The Humanist culture has also been around since the early years (e.g., query-by-example), but this culture too started growing mostly only from the 1990s. This culture draws inspiration from the fields of human computer interaction, programming languages, and cognitive science, but also parts of theoretical CS and computer systems. This culture puts humans that work with data at the center of the research world. Topics in which this culture is well-represented are new abstract programming/computation models for managing/processing data, new query interfaces, interactive data exploration/analysis tools, and data visualization. Common elements in Humanist papers are terms such as usability and productivity, user studies, and even links to demo videos of the tools they build. Remarkably, many Humanist papers often overlap with one or more of the other cultures, which perhaps makes this culture the least dogmatic and most eclectic. Humanist papers often have multiple metrics, including quality, system runtime, and human effort (measured with "lines of code", human interaction time, interviews, etc.). Many Humanists also publish at CHI/UIST.

This culture seems to be growing, primarily by attracting people from the other cultures. Working on human-centered problems has a long history in this field, going back to Codd himself. Many researchers often seem to forget the fact that the relational model itself was created not to solve a "hard" theoretical problem, improve a system's performance, or design a heuristic algorithm, but rather to improve the productivity of humans that queried structured data.

Culture Wars and Extremism

By "wars," I mean the endless intellectual tussle for suzerainty over the research field. The most common way such wars cause damage is unfair negative evaluations of research papers because one is unable to see the merits of a paper from a different culture or a cultural hybrid. Many "extremist" Formalists and Heuristicists (and Formalist+Heuristicists) often dismiss many Systemist papers as "just engineering." Many extremist Formalists also dismiss many Heuristicist papers as "too ad hoc" or "too flaky." Many extremist Systemists dismiss many Formalist papers as "not practical" or "too mathy." Many extremist Formalists and Systemists (and Formalist+Systemists) dismiss many Humanist papers as "fluff" and "soft science." Many extremist Formalists, Heuristicists, and Systemists (and many cultural hybrids) also often dismiss many Systemist or Heuristicist papers that explore new important problems and propose initial solutions as "too straightforward" or "not novel" by conflating simplicity, an oft-exalted virtue in the real world, with a lack of novelty. Overall, one often ends up wrongly judging a fish by its ability to climb a tree. Sadly, such wars sometimes force researchers to add needlessly contrived content to papers just to appease such extremists.

Such tribal culture wars and extremism, whether deliberate or not, detract from fair and honest critiques that actually help advance the science. Personally, I find such wars ridiculous. So, please allow me to amuse myself (and hopefully, you) by ridiculing the extremist bigots of each culture that ridiculously diss other cultures and glorify only their own using the provocative meme of "X as seen by Y" (see this one about programmer wars first, if you do not know this meme). Since a 16x16 matrix is too big for me to construct, I restrict myself to the canonical 4x4. :) Hopefully, this will help students/researchers realize that they are not alone in getting caustic comments. I also hope this will cause people to think twice before engaging in such ridiculous wars themselves in the future. Behold, I present to you the culture wars of data management!

(Optional) Explanatory Caption (might be painfully obvious for some). In row-major order from the top-left. First row: Einstein (geniuses, of course), FSM (mushy false gods), Crashed truck (what a disaster!), and Big fat hacker/engineer. Second row: Lucius Malfoy (snooty pure-blood evil wizards destined to be defeated), Justice League (superheroes saving humans), Awkward nerd, and Zealots (close-minded bigots serving the devil). Third row: Kung Fu Panda 3 (googly-eyed admiration), Pretend-Superman, Star Trek (the future beckons), and Airplane pilots (so many moving parts!). Fourth row: Irrelevant contrived junk peddlers, Kids with fun toys (so cute!), Calvin-and-Hobbes's games (so naive!), and The One (the prophetic savior).


Making matters worse are "civil wars" within cultures, especially the Systemists. Some Systemists are so obsessed with relational DBMSs that they pooh pooh any new data systems. Such insularity has caused much grief in the last decade, especially due to the rise of "Big Data" systems (think MapReduce/Hadoop or Spark) and "NoSQL" systems (think BigTable) from the distributed systems field. Of course, most Systemists acknowledge their mistakes and change their minds over time, but not without causing serious damage to the field. There is also a mini civil war among the Formalists between the logic and discrete math-oriented sub-culture and statistics/ML theory-oriented sub-culture. With so many culture wars going on, I fear research on "data management for ML"/"ML systems," which is the area my own research focuses on, will be driven away from SIGMOD/VLDB. This area is increasingly considered important for the wider CS landscape and thus, for the data management community too. I will be raising these issues (among others) at a panel discussion at the DEEM Workshop at SIGMOD 2018. Rest assured, I will return to blog about how that goes.

Should the Four Cultures Stick Together or Break Up?

The data management/database/data systems community is not a "monoculture"; it never was and it never will be (almost surely). As a vertical slice of CS, it is "multicultural" and will always have high intellectual diversity. The four cultures may be irreconcilable, but they are complementary and can co-exist. The benefits of inter-cultural tolerance, cultural hybridization, and trans-cultural work are clear: cross-pollination of problems and ideas, trans-cultural partnerships and collaboration, infusion of ideas from each culture's favored non-data management fields, export of ideas from one culture via another to another CS field or non-CS disciplines, and so on. This sort of inter-cultural amity and partnership was/is the norm at the Database Group of the University of Wisconsin-Madison, where I went to graduate school, and the Database Lab of the University of California, San Diego, where I am on the faculty now, and many other database groups. There is a long tradition of research cultural hybridization and trans-cultural research that is practiced and even celebrated by many researchers, senior and junior alike.

Yet, to claim such culture wars do not exist is to be the proverbial ostrich that buries its head in the sand, while pretending we can all just sing kumbaya, pat ourselves on our backs, and get along with stiff upper lips is to be utterly naive. We have to acknowledge and accommodate the vast cultural gaps amicably lest the centrifugal forces caused by unfair research evaluation and tribal cultural wars spin out of control. This form of multiculturalism is a source of strength for SIGMOD/VLDB and sets them apart from related communities such as STOC, OSDI, ISCA, SIGKDD, and CHI. Of course, it will also defeat the point of inter-cultural amity if everyone is expected to work across cultures; it should be left to each researcher to pick a culture or cultural hybrid for their work based on their skills and taste.

Instead of trying to shame or bully any of the cultural groups into conformity to another culture, all groups should practice mutual respect. An analogy I can draw is with movie genres: one cannot coerce a person that only likes arthouse dramas to like big-budget blockbusters or vice versa. All 4 cultures and all cultural hybrids bring something different and valuable to the table of data management research. It will be nothing short of a catastrophe for SIGMOD/VLDB if the Systemists (or indeed, any other group) leave. While such a pyrrhic war has been averted for now, I fear it might turn in to a Cold War that poisons the field further. Unfortunately, the current peer review processes of SIGMOD/VLDB are broken and are amplifying the centrifugal forces. Thankfully, they are aware of this issue and working to fix them soon. I am reasonably confident they will course correct but I doubt it will happen quickly or easily.

All that said, I could be wrong and perhaps the cultures are better off breaking up. As such, attending a conference is no longer needed for following new research thanks to the Web. Cutting-edge "data management" work is no longer published at SIGMOD/VLDB only; NSDI, OSDI, SIGKDD, NIPS, and more share the pie. Inter-cultural research exchange can happen across communities too. There are precedents for both area-based and culture-based splits/spinoffs: PODS (for database theory), CIDR (for "visiony" data systems), SoCC (for cloud computing), and SysML (for ML systems). Who can be 100% sure it is "bad" to have more splits/spinoffs?

Even if SIGMOD/VLDB remain multicultural, should they become more "federated" instead of "unitary" in terms of the research tracks? It is clearly unfair to have a random SODA (or SIGKDD) reviewer evaluate a random OSDI (or CHI) paper and vice versa--and yet, this is roughly the kind of unfairness caused by the culture wars that continue unabated within the SIGMOD/VLDB community. Is it not a form of "emotional abuse" of students to subject them to such toxic subjective culture wars instead of a resolute focus on the objective (de)merits of ideas? Does the SIGMOD/VLDB community really want to cut such a sorry figure against NSDI/OSDI/SOSP or NIPS/ICML or CHI or other communities that compete for bright students? I think the SIGMOD/VLDB community should openly and honestly tackle such contentious questions without acting like this or this. But I do not know the best avenue for spurring such conversations: a short workshop, a townhall or panel discussion at SIGMOD/VLDB, or something else, who knows?

Concluding Remarks

Scientists are people. It is crucial to practice objectivity and apply solid quality filters for research evaluation. But it is also crucial to be aware of one's own blind spots. Unknown unknowns are also impossible to fathom for anyone, no matter how knowledgeable--none of the blind men can say anything about parts of the elephant they cannot reach. This is all the more true for scientists at the cutting edge of the ever-expanding universe of knowledge. Also crucial is respecting subjective differences on intellectual practice, since the leap from objective facts to judgmental interpretation is far too often a leap of faith guided by subjective experience--even if two of the blind men grasp the same tail fur, one might call it soft on his hand and the other, coarse. Perhaps I am being a naive idealist, but I think it is imperative that we all inject ourselves with large doses of scientific humility, curiosity, civility, empathy, and magnanimity. Watching this video regularly might help! :) Unless there is credible verifiable evidence to the contrary, every scientist must be willing to admit that they could be wrong. Every one of us is (intellectually speaking) "blind" in some way, and we will remain blind no matter how much we learn. But we can all be less blind by talking with others that have an earnest world-view that is different from ours, not talking down to them.


PS: I do not claim that these 4 cultures are the only ones in data management. I am sure as the field keeps growing, we might see the addition of new cultures or the reorganization of existing ones. Perhaps I will write another post on this topic 20 years from now! If you have any thoughts on this topic or this post, please do leave your comments below.

ACKs: Special thanks to Julian McAuley for being a sounding board for this post. I also thank the following people (in alphabetical order of surnames) for giving me feedback on this post, including suggestions that strengthened it: Peter Bailis, Spyros Blanas, Vijay Chidambaram, Joe Hellerstein, Paris Koutris, Sam Madden, Yannis Papakonstantinou, Jignesh Patel, Christopher Re, Victor Vianu, Eugene Wu, and a few others that did not want to be acknowledged. My ACKs do not necessarily mean they endorse any of the opinions expressed in this post, which are my own.

---

Reactions (Contributed Paragraphs)

The post elicited a wide range of reactions from data management researchers. Some researchers have kindly contributed a paragraph to register their reactions here, both with and without attribution. I expect to have more contributions in due course. If you do not want to comment publicly, feel free to email me your reaction so that I can add it here as an anonymous contribution. I hope all this contributes to stirring more conversations on this important topic.

Jignesh Patel:
Interesting! Note Systemists don't abhor theory, they abhor theory for the sake of theory. Classic systems paper have some formalism, but just what is required to understand the system implication (think of the classic Gray locking paper) I don't think Systemists wage wars. Like true system people, they are quick to identify a practical problems, and speak up. Having said all that, you do bring out an important problem of tribal wars that is killing the community. Also, we actually have a really nice database systems community where the senior folks actually take criticism quite well. Arguments are welcome! Don't know for sure about the theory folks, but I think they are pretty open-minded too. Overall, we have a pretty nice community, and largely the right things happen in the long-run. Ok--I'm tenured and I'm not as pessimistic as other are perhaps as a result?

Anonymous:
I liked how you have broken up the community into 4 tribes. I agree with most of it. However, on a higher-level feedback, here's what I think: I feel like this is perhaps triggered by bad/unfair reviews which every one of us has dealt with. However, I personally see the poor reviewing issue (i.e., lack of tolerance) only a symptom of the bigger problem which probably has a lot less with "culture" and has more to do with a "mafia" mentality in the community. Awards/Recognitions/Opportunities are not distributed based on merits. Rather, advisors/friends look out for their own advisees/friends. For example, the more famous the advisor, the more opportunities for their students. That means most everyone else is simply an "academic orphan" in the community. Not all but the majority of senior people in the community who are in control of how recognitions/opportunities are distributed do NOT act on what they preach. In public, most of them talk about the importance of "impactful work" but when you read their own letters they are doing nothing more than "bean counting" when it comes to assessing impact. The traditional data management community is rapidly losing its relevance, and as such, everyone is trying to come up with a definition of what's a fake problem and what's real/worthy. How does this relate to low-quality reviews? The data management community is a hostile environment due to its contentious and unfair interworkings. In a hostile environment individuals aren't acting rationally, let alone fairly. For example, even PC chairs and area chairs are not always people who are deemed by the majority of the community as reasonable or even insightful.