feedbacklog.typ

#import "/layout/feedbacklog_template.typ": *
#import "/metadata.typ": *
#import "/utils/feedback.typ": *

#set document(title: titleEnglish, author: author)

#show: feedbacklog.with(
  titleEnglish: titleEnglish,
  examiner: examiner,
  supervisors: supervisors,
  author: author,
  presentationDate: presentationDate,
  feedbacklogSubmissionDate: feedbacklogSubmissionDate,
)

#feedback(
  feedback: "Is the resource pool, as the main contribution, independent of the programming exercise language?",
  response: "Yes, the resource pool is not directly tied to specific programming exercises.",
)

#feedback(
  feedback: "The pool does not load the JDK etc., since prebuilt IDE images are used instead. However, the pool still thinks of images differently, so isn't there a dependency after all?",
  response: "There is a dependency on images for a specific language, but those images are shared by many programming exercises.",
)

#feedback(
  feedback: "Was the benchmark done with only one specific blueprint?",
  response: "Yes, the benchmark was only done with one language configuration, Java 17 to be specific.",
)

#feedback(
  feedback: "Can you explain how the resource pooling works? What happens if there are many different configurations?",
  response: "Prewarming happens separately per language image, so there is a pool for each image. Different pools are unrelated to each other.",
)

#feedback(
  feedback: "How many prewarmed resources were there for the benchmark?",
  response: "For the benchmark with 50 concurrent users, there were 50 prewarmed resources. If demand exceeds that, the system falls back to lazy provisioning.",
)

#feedback(
  feedback: "Do you think it is realistic to have 50 prewarmed sessions?",
  response: "Yes, the system can definitely support over 100 without any issues, so this is realistic.",
)

#feedback(
  feedback: "Are prewarmed resources created at the same time? Does prewarming happen automatically in the background?",
  response: "Predictive scaling would ideally determine the size of the pool automatically, but currently this is done manually.",
)

#feedback(
  feedback: "How many resources does one session need?",
  response: "About 0.0006 of CPU and 265 MiB of RAM.",
)

#feedback(
  feedback: "So the prewarmed resources basically sleep until needed - is waking them up faster?",
  response: "Yes, since it is just an HTTP service that isn't being used, waking it up is fast.",
)

#feedback(
  feedback: "The presentation was good overall, with mostly dynamic behavior and a good visual representation. Matthias really liked the thesis and the work done.",
  response: "",
)

#feedback(
  feedback: "(About future ideas) How exactly do you determine how many resources you want to provision?",
  response: "Deterministic data is used first, such as exam schedules, since there is no problem having too many resources. As a further optimization, one could look into more advanced machine learning models to predict demand.",
)

#feedback(
  feedback: "(About future ideas) Is this deterministic data provided by the instructors?",
  response: "No, the data is taken directly from Artemis, without any manual intervention from the instructors. The predictive scaling should function without human intervention.",
)