You are here

Feed aggregator

Four success factors for building your AI business journey

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Manish Goyal shows you how to best unlock the value of enterprise AI.

Continue reading Four success factors for building your AI business journey.

Categories: Technology

AI and security: Lessons, challenges, and future directions

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Dawn Song explains how AI and deep learning can enable better security and how security can enable better AI.

Continue reading AI and security: Lessons, challenges, and future directions.

Categories: Technology

Connected arms

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Joseph Sirosh tells an intriguing story about AI-infused prosthetics that are able to see, grip, and feel.

Continue reading Connected arms.

Categories: Technology

Customized ML for the enterprise

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Levent Besik explains how enterprises can stay ahead of the game with customized machine learning.

Continue reading Customized ML for the enterprise.

Categories: Technology

The breadth of AI applications: The ongoing expansion

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Peter Norvig says one of the most exciting aspects of AI is the diversity of applications in fields far astray from the original breakthrough areas.

Continue reading The breadth of AI applications: The ongoing expansion.

Categories: Technology

Accelerating AI on Xeon through SW optimization

O'Reilly Radar - Fri, 2018/09/07 - 13:00

Huma Abidi discusses the importance of optimization to deep learning frameworks.

Continue reading Accelerating AI on Xeon through SW optimization.

Categories: Technology

A new golden age for computer architecture

O'Reilly Radar - Fri, 2018/09/07 - 13:00

David Patterson explains why he expects an outpouring of co-designed ML-specific chips and supercomputers.

Continue reading A new golden age for computer architecture.

Categories: Technology

10 talks to look for at the 2018 O'Reilly Software Architecture Conference in London

O'Reilly Radar - Fri, 2018/09/07 - 08:10

From chaos architecture to event streaming to leading teams, the O'Reilly Software Architecture Conference offers a unique depth and breadth of content.

We received more than 200 abstracts for talks for the 2018 O'Reilly Software Architecture Conference in London—on both expected and surprising topics. We continue to see strong interest in microservices and its related ecosystem, including topics like DevOps and tools like Kubernetes. The quality of the abstracts led to a stellar lineup of speakers, talks, and keynotes.

Two of the outstanding features of the O'Reilly Software Architecture Conference are the depth and breadth of our content. While most conferences have a single software architecture track, our whole conference revolves around software architecture. That means we can go much deeper, covering topics that would be too rarefied for other conferences. That also means we can spread out, tackling subjects critical to success as an architect (like soft skills) but too broad for most developers conferences. To showcase our depth and breadth, I've chosen a few sessions to highlight, illustrating the astounding variety of topics and perspectives on display this year.

Introduction to Chaos Architecture: Gaining from Learning Loops and System Weaknesses, by Russ Miles, ChaosIQ.io

Chaos engineering and the attendant architectural concerns is a red-hot topic—we had a lot of interesting talks at the Software Architecture Conference in New York this year, too. In the talk linked here, Russ Miles asks the pointed question: "What happens when (not if) something in your system breaks? How will you handle it?" Pioneered by Netflix, chaos engineering represents a great example of continuing innovation in software engineering practices.

Technology Strategy Patterns for Architects, by Eben Hewitt, Sabre

This is cheating, as both this and the previous talks happen at the same time, but I can't not mention both! I had the privilege to review Hewitt's upcoming O'Reilly book by the same title. A great example of our breadth, this talk exposes the uninitiated to the seemingly arcane world and nomenclature of strategy consultants. Architects must often participate in these types of meetings, so understanding the building blocks and approaches of these consultants allows architects to participate and contribute. A highly recommended talk in an area almost never covered at technology conferences.

7 Years of DDD: Tackling Complexity in Large-Scale Marketing Systems, by Vladik Khononov, Naxex

We get a huge number of excellent case studies at the O'Reilly Software Architecture Conference, which we like because it gives attendees insight into real-world problems and solutions; this one is an outstanding example. Many organizations think about embracing the domain-driven design (DDD) philosophy but wonder about the long game—where will we be seven years from now? This talk provides useful perspective of the long-term implications, problems, and solutions inherent in fully embracing a technique like DDD.

Observable Microservices, by Maria Gomez, ThoughtWorks

Like peanut butter and chocolate, monitoring and microservices are a natural combination. This talk delves into the nuances of building microservices that incorporate sophisticated monitoring and fitness functions for a better understanding of the runtime characteristics of your system.

Beyond the Technical—Succeed at Leading a Software Architecture Team, by Maggie Carroll, Ausley.us

Many architects are shocked when they realize how many non-technical skills are required to be successful at their jobs, and team leadership tops the list. This talk covers critical skills and perspectives for architects to embrace to help team building.

Architecting for Data-Driven Reliability, by Yaniv Aknin, Google Cloud

Data is often ignored in architecture talks as a messy inconvenience, but the real world doesn't allow that luxury. As architectures become more distributed, data reliability becomes a critical concern. This talk covers how to design an architecture to ensure modern capabilities while still maintaining old-school reliability.

Event Streaming as a Source of Truth, by Benjamin Stopford

As teams build more sophisticated distributed systems, they sometimes move to event streams rather than databases as the source of truth. However, architects must deal with numerous issues and considerations when making such a fundamental shift. This talk covers the pros and cons, along with some best practices and warnings.

Distributed Systems Are a UX Problem, by Tyler Treat, Real Kinetic

No matter how distributed many architectures become, they still re-unify at a monolithic user interface. Many architects struggle reconciling a highly distributed microservices architecture with a monolithic user interface; this talk covers many of the issues and offers some solutions.

Sundhed.dk's Journey From Monolith to GDPR-Compliant Microservices, by Tobias Uldall-Espersen and Thomas Krogsgaard Holme, Sundhed.dk

This journey is another great case study. Many attendees find themselves on the same journey of restructuring a monolithic application to microservices, and many have been recently blindsided by GDPR. This talk offers great real-world insight into how to migrate and mitigate architectural concerns simultaneously.

Using Continuous Delivery with Machine Learning to Tackle Fraud, by Sarah LeBlanc and Hany Elemary, ThoughtWorks

This talk is a great mashup of ideas that could only find a home at a software architecture conference! Yet another case study-based talk, it covers a real-world concern that many companies struggle with by combining state-of-the-art engineering practices, machine learning, and security, and tying it all together with architecture.

As the O'Reilly Software Architecture Conference has grown over the last couple of years, the quality of talks continues to rise. The big problem is choosing between what’s on offer in every time slot.

Continue reading 10 talks to look for at the 2018 O'Reilly Software Architecture Conference in London.

Categories: Technology

Four short links: 7 September 2018

O'Reilly Radar - Fri, 2018/09/07 - 05:00

Quantifying Facebook, Deep Learning IDE, REPL + Debugger, and RPC Library

  1. Unveiling and Quantifying Facebook Exploitation of Sensitive Personal Data for Advertising Purposes -- This paper quantifies the portion of Facebook users in the European Union (EU) who were labeled with interests linked to potentially sensitive personal data in the period prior to when GDPR went into effect. The results of our study suggest that Facebook labels 73% of EU users with potential sensitive interests. This corresponds to 40% of the overall EU population. We also estimate that a malicious third party could unveil the identity of Facebook users who have been assigned a potentially sensitive interest at a cost as low as €0.015 per user. Finally, we propose and implement a web browser extension to inform Facebook users of the potentially sensitive interests Facebook has assigned them. (via Morning Paper)
  2. Subgraphs -- a deep learning IDE.
  3. REPLugger: REPL + Debugger -- My belief is that providing tools to augment programmer understanding is one of the most important interventions we can make. Me, too.
  4. brpc -- Baidu's RPC library, with 1,000,000+ instances (not counting clients) and thousands of kinds of services.

Continue reading Four short links: 7 September 2018.

Categories: Technology

Using machine learning in workload automation

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Akhilesh Tripathi shows you how to use machine learning to identify root causes of problems in minutes instead of hours or days.

Continue reading Using machine learning in workload automation.

Categories: Technology

China: AI superpower

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Kai-Fu Lee outlines the factors that enabled China's rapid ascension in AI.

Continue reading China: AI superpower.

Categories: Technology

AI foundations: What shapes the AI that’s shaping our world?

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Meredith Whittaker says the benefits of AI will only come if we have a clear-eyed perspective on its dark side.

Continue reading AI foundations: What shapes the AI that’s shaping our world?.

Categories: Technology

Fireside chat with Tim O'Reilly and Kai-Fu Lee

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Tim O'Reilly and Kai-Fu Lee discuss differences in how China and the U.S. approach AI and why AI might give humanity larger purpose.

Continue reading Fireside chat with Tim O'Reilly and Kai-Fu Lee.

Categories: Technology

Raising AI to benefit business and society

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Kishore Durg explains why deploying AI requires raising it to act as a responsible representative of the business and a contributing member of society.

Continue reading Raising AI to benefit business and society.

Categories: Technology

Highlights from the Artificial Intelligence Conference in San Francisco 2018

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Watch highlights from expert talks covering artificial intelligence, machine learning, security, and more.

People from across the AI world came together in San Francisco for the Artificial Intelligence Conference. Below you'll find links to highlights from the event.

China: AI superpower

Kai-Fu Lee outlines the factors that enabled China's rapid ascension in AI.

Fireside chat with Tim O'Reilly and Kai-Fu Lee

Tim O'Reilly and Kai-Fu Lee discuss differences in how China and the U.S. approach AI and why AI might give humanity larger purpose.

AI foundations: What shapes the AI that’s shaping our world?

Meredith Whittaker says the benefits of AI will only come if we have a clear-eyed perspective on its dark side.

Using machine learning in workload automation

Akhilesh Tripathi shows you how to use machine learning to identify root causes of problems in minutes instead of hours or days.

AI at scale at Coinbase

Soups Ranjan describes the machine learning system that Coinbase built to detect potential fraud and fake identities.

OpenAI and the path toward safe AGI

Greg Brockman discusses OpenAI's recent advancements and their implications for how we should plan for creating safe artificial general intelligence (AGI).

Raising AI to benefit business and society

Kishore Durg explains why deploying AI requires raising it to act as a responsible representative of the business and a contributing member of society.

Beyond hype: AI in the real world

Julie Shin Choi reviews real-world customer use cases that take AI from theory to reality.

Machine learning in the cloud

Hagay Lupesko explores key trends in machine learning, the importance of designing models for scale, and the impact that machine learning innovation has had on startups and enterprises alike.

Customized ML for the enterprise

Levent Besik explains how enterprises can stay ahead of the game with customized machine learning.

Accelerating AI on Xeon through SW optimization

Huma Abidi discusses the importance of optimization to deep learning frameworks.

AI and security: Lessons, challenges, and future directions

Dawn Song explains how AI and deep learning can enable better security and how security can enable better AI.

Four success factors for building your AI business journey

Manish Goyal shows you how to best unlock the value of enterprise AI.

The breadth of AI applications: The ongoing expansion

Peter Norvig says one of the most exciting aspects of AI is the diversity of applications in fields far astray from the original breakthrough areas.

Connected arms

Joseph Sirosh tells an intriguing story about AI-infused prosthetics that are able to see, grip, and feel.

A new golden age for computer architecture

David Patterson explains why he expects an outpouring of co-designed ML-specific chips and supercomputers.

Continue reading Highlights from the Artificial Intelligence Conference in San Francisco 2018.

Categories: Technology

Beyond hype: AI in the real world

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Julie Shin Choi reviews real-world customer use cases that take AI from theory to reality.

Continue reading Beyond hype: AI in the real world.

Categories: Technology

AI at scale at Coinbase

O'Reilly Radar - Thu, 2018/09/06 - 13:00

Soups Ranjan describes the machine learning system that Coinbase built to detect potential fraud and fake identities.

Continue reading AI at scale at Coinbase.

Categories: Technology

How network professionals deal with attacks and disruptions

O'Reilly Radar - Thu, 2018/09/06 - 08:20

A new survey highlights concerns from network and cloud administrators, and reveals their coping strategies.

Reliability and response time are always pressing concerns for internet services. We’ve known for a long time that people on desktop or laptop computers have scant patience for slow websites, and the growing move to mobile devices makes the demands on internet services even worse.

O'Reilly Media and Oracle Dyn teamed up this year to survey network operators about where their resilience problems lie, what they're doing to avoid these problems, and what it's like to work in the field of network operations. The survey paid particular attention to issues of working in third-party cloud services, but covered a wide range of other problems in networking as well. Among the findings, we were surprised to hear the chief complaint of respondents in regard to resilience: their own ISP! We also uncovered a high concern over staff burnout, and disparities in the training of operators and in the handling of incidents.

This article starts with a brief description of the survey and who responded. Then we’ll get into the most interesting of the findings. We’ll end with some speculations about automation.

Who responded to the survey?

We collected answers to our survey from 621 network professionals with a variety of job descriptions (see Figure 1). Respondents work in many different industries, although IT services and software dominate. Organizational size was lopsided at the edges of the spectrum (Figure 2), with the largest fraction of respondents employing fewer than 200 people, and the next largest fraction having 10,000 or more. Small percentages represented in-between company sizes.

Figure 1. Industries where respondents work. Image: O'Reilly. Figure 2. Number of employees in organizations where respondents work. Image: O'Reilly.

The majority of respondents taking the survey (459, or 74%) reported working in the cloud. These respondents were asked further questions about what resilience problems they experienced and what measures they took to deal with them (Figure 3). The most common use for the cloud­ was apps/services (81% of cloud users) and development/testing (71%). Historically, we know that many organizations started their cloud use with development and testing before taking the leap of putting their production systems there. As one might expect, the cloud is also popular for web hosting, scaling compute power or storage, and backups.

Figure 3. How respondents use the cloud. Image: O'Reilly. The biggest single problem: Your own ISP

The most surprising result from our survey may be the answer to what respondents said was the leading cause of network disruptions: ISP reliability issues (Figure 4), cited by 44% of respondents. If you add up the various malicious attacks reported, they collectively surpass ISP problems. But given the prevalence of ISP problems, it makes sense that some kind of redundancy is the most popular way to ensure resilience: 39% of respondents use load balancing and 33% use multiple failover sites. Smaller but still significant numbers run a second DNS service, use multiple cloud providers, and use multiple ISPs (Figure 5).

Figure 4. Disruptions experienced by cloud users. Image: O'Reilly. Figure 5. How cloud users avoid disruptions. Image: O'Reilly. The cloud has not overtaken the landscape

The large percentage of respondents who over-provision resources (19%) suggests that on-premises deployments are still common, even among sites using the cloud as well. Overprovisioning is a common practice in on-premises deployments, and is usually rendered unnecessary in the cloud by autoscaling. And it's worth noting that 29% of respondents rely entirely on the cloud provider for resilience (Figure 5, above). A variety of SLAs and incident response remedies are in use (Figure 6).

Figure 6. Use of SLAs and responses to incidents. Image: O'Reilly. You’re not paranoid

Or you might as well be, because they’re out to get you. Attack vectors suggest that sites taking pre­emptive steps to mitigate the impact of attacks are justified. For example, among responses indicating the use of DDoS mitigation tools, 51% (85 of 166 respondents, Figure 7) also indicated that their organization actually had a DoS/DDoS attack in the last 12 months. This is nearly twice that of the overall share reporting DoS/DDoS attacks had occurred overall (121 of 459 respondents).

Figure 7. Disruptions experienced, broken down by measures taken to reduce disruptions. Image: O'Reilly.

A few other correlations turn up between the disruptions experienced and the measures taken for resilience:

  • Using multiple failover sites that share traffic is a popular way to ensure resilience, but it is particularly popular among those who experience DoS attacks or IP route hijacks.
  • Companies that experience those attacks (especially DoS attacks) are more likely to have a secondary DNS service. So are firms having problems with ISP reliability.
  • Spam and phishing were cited as problems by a large number of sites in general (around one-third) but somewhat more by sites that perform black-holing or sink-holing.

Disruptions don’t happen often, but they have to be expected. Each organization was likely to experience just one or two disruptions, but as we explained at the beginning, even a single disruption can be a serious concern (just remember the famous Netflix Christmas outage). A lucky or highly disciplined 20% of respondents reported no disruptions or interruptions during the past year.

Neither organizational size nor industry type made a difference in the problems faced by organizations or the responses they implemented. Similarly, no particular worry dominated responses to the survey. Everything was considered important by a substantial fraction of respondents, although none of them were cited by a majority. And all cloud users employed similar techniques to ensure resilience, regardless of their purpose for using the cloud.

Solutions are at hand—but not in use

It is interesting how many organizations fail to use certain preemptive tools that one might think are universal (Figure 8). For instance, among respondents who oversee the use of cloud services at their organization, more than one-third (37%) said they don’t use firewalls. Even more (42%) fail to use health monitoring. Similarly, DDoS mitigation is used by only 36% of respondents using the cloud. Perhaps the others don’t expect DDoS mitigation to be effective, or anticipate that their services can just wait out a DDoS attack. Or perhaps these respondents rely on the cloud provider to run these services, so some of the services might, in fact, be in place despite the responses we received.

Figure 8. Use of pre-emptive tools to avoid disruptions in the cloud. Image: O'Reilly.

This section wraps up our findings about outages, attacks, and technologies. Now we turn to the human element of network monitoring. Are there enough educated staff in this field? How are companies educating their staffs, and how does that affect responses to network disruptions?

Good network operators are at a premium

In 2015, network engineering was listed as one of the best career choices in computing. The article making that claim does not explain its criteria, but it apparently counts just the demand for such positions: 105,000 openings that year. Job conditions are a different matter, and our survey responses suggest that not all is well in the working conditions for network operators.

Respondents to the survey tend to move around. Fewer than 35% had been in their current jobs for more than five years. We didn't ask how long they had been working in total as network operators. But the greatest single concern respondents expressed about their staff was lack of experience (53%) and insufficient training (46%), both of which suggest that a lot of operators are fairly new to the field. Most organizations are willing to train new operators, either through formal training programs (38%) or informal mentoring (38%).

A substantial number of organizations (29%) hire only experienced people, a luxury that may elevate salaries throughout the field. One can expect people in this field to be highly employable, and therefore to move around a lot in search of more money or better work environments.

In this regard, it's significant that burnout was a major concern among respondents (46%), along with turnover (37%). This should not be surprising because in most institutions, network operations involve constant pressure. And no one likes those 3 a.m. wake-up calls.

Overall, respondents were concerned less about losing employees (turnover) and more about ensuring current employees remained satisfied in their jobs (i.e., that gaps in experience or training are addressed and that employees don’t burn out).

Sites handle outages and attacks differently

Many common forms of analysis and recovery are performed after an attack (Figure 9). The methods organizations use to respond to attacks or to ensure resilience are sometimes correlated with their efforts at skills development, and perhaps even influence staff burnout.

Figure 9. How organizations analyze and recover from disruptions. Image: O'Reilly.

Some examples of different responses include:

  • Among respondents who indicate that they rely completely on cloud service providers (135 respondents), only a little more than one-­third indicated that their organization protects against burnout through either advanced training or career advancement opportunities (35% and 36%, respectively). These percentages are considerably lower than the use of training or other tools to prevent burnout by respondents who employ load balancing across sites or the use of multiple failover sites.
  • The approach used for training network ops teams (Figure 10) may influence how incidents are handled in postmortems. For example, just 37% of those who only hire experienced people indicated they document “lessons learned” to ensure preventable incidents don’t occur.
Figure 10. How network operators are trained to handle resilience. Image: O'Reilly.
  • Among those who train their staff through internal programs, reliance on vendors, or employee "shadowing," around half also reported their postmortems documented lessons learned (48%, 52%, and 58%, respectively).
  • Rewards and recognition are the slightly preferred approach to addressing burnout (Figure 11). Notably, though, this approach does not appear to be considered enough for the majority. Of the 223 respondents who selected rewards and recognition (49% of total respondents), nearly two-thirds (65%, 145 respondents) also selected career advancement opportunities or advanced training to protect operations staff against burnout.
Figure 11. How organizations attempt to prevent burnout. Image: O'Reilly. Automation: Is there a better way?

The survey did not ask what forms of automation organizations use, but these are probably lacking. Despite the current enthusiasm for DevOps in development, the 2017 Puppet "State of DevOps Report" found that about 40% of survey respondents still do manual configuration management and deployment.

Thus, we can surmise that a large number of network operators have to suffer through repetitive recovery tasks under high pressure, along with being tied to their pagers. Even among the 60% of organizations that do some automation, it is probably incomplete. As tools become more widespread and better understood—and especially as cloud providers make them simple to deploy—we can look forward to less burnout and turnover. In summary, given how widespread burnout appears to be, organizations should perhaps make automation a priority.

Conclusion

Overall, our survey suggests that organizations are surviving network outages or reliability problems pretty well. Concerns about resilience, and the measures taken to address them, are fairly consistent across industries, organizational size, and the attacks or failures encountered.

ISP failures are a major concern, vying in importance with malicious attacks. In response to both concerns, sites employ many forms of redundancy, ranging from over-provisioning resources to using multiple cloud providers and DNS services. On the other hand, many respondents are happy sticking their services in the cloud and allowing the cloud provider to deal with reliability.

Warning flags crop up in the treatment of network operators and their job satisfaction. The pressure to hire and retain operators can be addressed by more training or by reducing the need for such staff through automation. Measures that seem to be put in place for the benefit of junior engineers, such as documenting incidents, may end up being healthy for the organization as a whole.

Perhaps this article will encourage more managers of network operations to take a closer look and expand their use of tools for preventing and mitigating problems. Currently, according to the results of the survey, such practices are not as widespread as one would think. We also suggest that operators put more effort and be more consistent in their postmortem handling of incidents, invest in training, and improve the jobs through modern automated practices to prevent burnout.

This post is a collaboration between O’Reilly and Oracle Dyn. See our statement of editorial independence.

Continue reading How network professionals deal with attacks and disruptions.

Categories: Technology

Four short links: 6 September 2018

O'Reilly Radar - Thu, 2018/09/06 - 03:55

BS in AI, Visual Exploration, Bad Predictions, and USB-C Development

  1. CMU's AI Bachelor's Degree -- ethics course mandatory, likewise seven humanities courses. Nice.
  2. GANlab -- interactive visualization of what's happening in a generative adversarial network, as well as an easy-to-read explanation.
  3. Errors, Insights, and Lessons of Famous AI Predictions -- These case studies illustrate several important principles, such as the general overconfidence of experts, the superiority of models over expert judgement, and the need for greater uncertainty in all types of predictions. The general reliability of expert judgement in AI timeline predictions is shown to be poor, a result that fits in with previous studies of expert competence.
  4. USB-C Explorer -- a development board with everything needed to start working with USB Type-C. It contains a USB-C port controller and Power Delivery PHY chip, a microcontroller, and several options for user interaction.

Continue reading Four short links: 6 September 2018.

Categories: Technology

Four short links: 5 September 2018

O'Reilly Radar - Wed, 2018/09/05 - 04:15

Atomic Receiver, Nerdery as AR, Open Access, and Journey Maps

  1. An Atomic Receiver for AM and FM Radio Communication -- lasers detect fluctuations in the outer shell of "Rydberg vapors" (a special form of Cesium) that are caused by radio waves. See also MIT Tech Review.
  2. Geology is Like AR for the Planet (Wired) -- looking at the planet through a geologic lens is something like strapping on an augmented-reality headset. It invites you, from your vantage point in the present, to summon up Earth’s deep past and far future—to see these parallel worlds with your own eyes, like digital overlays. All nerd-level expertise is awesome for this reason. Try going bar-hopping with a bar owner who can talk about fit-out costs, eyelines, liquor choices, branding, etc. Nothing is boring if you know enough about it. (via Dan Hon)
  3. Radical Open-Access Plan (Nature) -- Eleven research funders in Europe announce "Plan S" to make all scientific works free to read as soon as they are published.
  4. Journey Maps -- A journey map is a collection of customer research most recognizable by its timeline—a visual depiction of every touch point customers have with the product or business, laid out from left to right. [...] Seeing the journey visually helps reveal the emotional landscape of the customer, which helps the product, marketing, customer support, and analytics teams understand what users feel at each point and identify ways the team can improve the experience. Steps and advice on how to build them.

Continue reading Four short links: 5 September 2018.

Categories: Technology

Pages

Subscribe to LuftHans aggregator