Uber's business is highly real-time in nature. PBs of data is continuously being collected from the end users such as Uber drivers, riders, restaurants, eaters and so on everyday. There is a lot of valuable information to be processed and many decisions must be made in seconds for a variety of use cases such as customer incentives, fraud detection, machine learning model prediction. In addition, there is an increasing need to expose this ability to different user categories, including engineers, data scientists, executives and operations personnel which adds to the complexity. In this paper, we present the overall architecture of the real-time data infrastructure and identify three scaling challenges that we need to continuously address for each component in the architecture. At Uber, we heavily rely on open source technologies for the key areas of the infrastructure. On top of those open-source software, we add significant improvements and customizations to make the open-source solutions fit in Uber's environment and bridge the gaps to meet Uber's unique scale and requirements. We then highlight several important use cases and show their real-time solutions and tradeoffs. Finally, we reflect on the lessons we learned as we built, operated and scaled these systems.

This article presents a 32-GHz frequency modulated continuous wave (FMCW) modulator based on the phase-locked loop (PLL) with nested sub-PLL structure in a 65-nm CMOS process. With the sub-PLL, the low-pass effect in phase domain is realized, reducing the noise folding effect, quantization noise, and spurs due to the delta sigma modulator (DSM). To achieve good stability and phase noise performance, both the phase-domain model and the phase noise model are analyzed and simulated. Based on these models, the chirp linearity is discussed and simulated, which helps to determine the design parameters and verifies the linearity improvement. The measurement results illustrate that in fractional-N mode, the nested-PLL achieves the phase noise of -91 dBc/Hz at 1-MHz offset frequency and the fractional spurs of less than -54 dBc at 30.78-GHz output frequency. In FMCW mode, the proposed modulator achieves a triangular chirp with 1.08-2.16-GHz bandwidth at about 32-GHz center frequency. In addition, the measured root mean square (rms) frequency errors of 400 and 770 kHz are achieved with the ramp slopes of 1.08 GHz/93 μs and 2.16 GHz/93 μs, respectively. Measurement results prove the improvements of the phase noise and chirp linearity with the sub-PLL. Including all pads, the chip occupies a silicon area of 1.5 mm2, and consumes 62-mW dc power.

This paper presents a 28-GHz low phase noise voltage-controlled oscillator (VCO) in a 65-nm CMOS process for 5G communication applications. With the capacitive splitting and transformer feedback techniques, theoretical analysis and simulations are undertaken for the proposed oscillator, proving that its negative Gmcan be enhanced over the interested frequency range and reduced at the harmonic frequencies. Moreover, the proposed oscillator startup performance is analyzed, and its tank quality factor, transient and phase noise performance are simulated. With 4-bit switch capacitors and varactors, the proposed VCO achieves a tuning range from 25.7 to 29.7 GHz, consuming 10.8 mW from a 0.9-V supply voltage. With a wideband common-source buffer, the output signal power is around 1 dBm. At 26.5 GHz, the proposed VCO achieves low phase noise of -105.8 dBc/Hz at 1-MHz offset and -130 dBc/Hz at the 10-MHz offset, respectively.

While Ajax programming and the plethora of JavaScript component libraries enable high-quality Uls in web applications, integrating them with page data is laborious and error-prone as a developer has to handcode incremental modifications with trigger-based programming and manual coordination of data dependencies. The FORWARD web framework simplifies the development of Ajax applications through declarative, state-based templates. This declarative, data-centric approach is characterized by the principle of logical/physical independence, which the database community has often deployed successfully. It enables FORWARD to leverage database techniques, such as incremental view maintenance, updatable views, capability-based component wrappers and cost-based optimization to automate efficient live visualizations. We demonstrate an end-to-end system implementation, including a web-based IDE (itself built in FORWARD), academic and commercial applications built in FORWARD and a wide variety of JavaScript components supported by the declarative templates.

This paper studies the following problem: given (1) a query and (2) a set of sensitive records, find the subset of records "accessed" by the query. The notion of a query accessing a single record is adopted from prior work. There are several scenarios where the number of sensitive records is large (in the millions). The novel challenge addressed in this work is to develop a general-purpose solution for complex SQL that scales in the number of sensitive records. We propose efficient techniques that improves upon straightforward alternatives by orders of magnitude. Our empirical evaluation over the TPC-H benchmark data illustrates the benefits of our techniques.

While Ajax-based programming enables faster performance and higher interface quality over pure server-side programming, it is demanding and error prone as each action that partially updates the page requires custom, ad-hoc code. The problem is exacerbated by distributed programming between the browser and server, where the developer uses JavaScript to access the page state and Java/SQL for the database. The FORWARD framework simplifies the development of Ajax pages by treating them as rendered views, where the developer declares a view using an extension of SQL and page units, which map to the view and render the data in the browser. Such a declarative approach leads to significantly less code, as the framework automatically solves performance optimization problems that the developer would otherwise hand-code. Since pages are fueled by views, FORWARD leverages years of database research on incremental view maintenance by creating optimization techniques appropriately extended for the needs of pages (nesting, variability, ordering), thereby achieving performance comparable to hand-coded JavaScript/Java applications.

Searching an organization's document repositories for experts is a frequently faced problem in intranet information management. This paper proposes a candidate-centered model which is referred as Candidate Description Document (CDD)-based retrieval model. The expertise evidence about an expert candidate scattered over repositories is mined and aggregated automatically to form a profile called the candidate's CDD, which represents his knowledge. We present the model from its foundations through its logical development and argue in favor of this model for expert finding. We devise and compare the different strategies for exploring a variety of expertise evidence. The experiments on TREC enterprise corpora demonstrate that the CDD-based model achieves significant and consistent improvement on performance through comparative studies with non-CDD methods.

Searching an organization's document repositories for experts is a frequently occurred problem in intranet information management. A common method for finding experts in an organization is to use social networks - people are not isolated but connected by various kinds of associations. In organizations, people explicitly send email to one another thus social networks are likely to be contained in the patterns of communication. Moreover, in some web pages, the relationship among people is also recorded. In our approach we propose several strategies in discovering the associations among people from emails and web pages. Based on the social networks, we proposed an expertise propagation algorithm: from a ranked list of candidates according to their probability of being expert for a certain topic, we select a small set of the top ones as seed, and then use the social networks among the candidates to discover other potential experts. The experiments on TREC enterprise track show significant performance improvement with the algorithm.

Performance evaluation is an important issue in Web search engine researches. Traditional evaluation methods rely on much human efforts and are therefore quite time-consuming. With click-through data analysis, we proposed an automatic search engine performance evaluation method. This method generates navigational type query topics and answers automatically based on search users. querying and clicking behavior. Experimental results based on a commercial Chinese search engine's user logs show that the automatically method gets a similar evaluation result with traditional assessor-based ones.