Variance Packet: Fill & Download for Free

A day or two before the meeting the recruiters upload all of the hiring packets to the tracking system. The lead recruiter for the specific HC sends an email to the members with a link to the page with that week's candidates (typically about ten).In the 24 or 48 hours before the meeting the HC members read the packets. For each packet each HC member writes comments, makes a recommendation (hire/no-hire) and sets a score (letter grades - A+ down to C- as I recall). When writing comments it was fairly common to copy particularly good text from the interview feedback into the comment to emphasize a key point.At the HC meeting one of the recruiters projects the dashboard of the day's candidates on a screen. The scores, hidden until then, are revealed and the entire committee discusses each candidate in turn. The list is sorted by average score (highest to lowest) and variance (lowest to highest).A score with low variance means that the committee, reading the packets independently, settled on a consensus view. My memory is that we had low variance for half to two-thirds of the candidates each week and were able to dispose of them with a clear decision quickly.When there was either a relatively low score or a lot of divergent views we would discuss the candidate. The facilitator would typically go around the table asking for comments or color from each HC member. We would also expand the detailed comments written by each of the HC members and review one anothers' written notes.If we could come to consensus, we did. If we failed we had several possible ways forward.We could ask for an extra interview or two to focus on specific topics that we felt that the interviewers had not adequately covered (often we would have an HC member volunteer to conduct one of these ... this was a high-stakes last-recourse approach, so we understood that the HC itself should be called on to help resolve uncertainty)We could reject the candidate - in general our strategy was "if you can not decide, the answer is no." We viewed this as a failure of our HC and tried to get to a decision.Because extra interviews were expensive in time and effort, we would try to get to a decision with the information available.We would often accommodate observers. These would be engineers previously trained as interviewers and with some experience with interviewing.About once a quarter we would review our statistics and look at how often our decisions were upheld or overturned by the next-more-senior committee (our decisions flowed up to a committee of VPs and Directors that could affirm or reverse our decisions and adjust the scores). We also looked at broad statistics, though I don't remember precisely what.

TCP estimates RTT using an Exponential Weighted Moving Average (https://en.m.wikipedia.org/wiki/Moving_average#Exponential_moving_average (sorry I can't figure out how to format hyperlink on mobile)) based on the observed RTT during transmission.Basically, when TCP sends a packet, it starts a timer, which will time out when it reaches the computed TimeoutInterval value. For simplicity, only one timer is used, despite the fact that several segments may be in transit at once (this is called pipelining).For any given packet, there is a sequence number. When the TCP sender receives an ACK for the packet, it stops the timer; the elapsed time is saved as SampleRTT, and provides a snapshot of network conditions. As this value can fluctuate with time, the EWMA is computed using the previous EstimatedRTT and the newly acquired SampleRTT by the equation:EstimatedRTT = 0.875 * EstimatedRTT + 0.125 * SampleRTTFor an explanation of the values used, consult RFC 6298. The use of an EWMA means that the relative importance of a given sample falls off exponentially as more are added. In English, more recent samples are considered more important because the represent a more recent estimate.In addition to the average, the RTT variability or deviation is estimated byDevRTT = 0.75 * DevRTT + 0.25 * |SampleRTT - EstimatedRTT|Given these averages, we need to decide on a timeout to set. Obviously the timeout should be at least EstimatedRTT, with an additional margin for variance. If DevRTT is small, the RTT is fairly constant, so the margin can be small. If the variance is higher, we should allow a larger margin for error. ThereforeTimeoutInterval = EstimatedRTT + 4 * DevRTTRFC 6298 recommends an initial TimeoutInterval of one second. When a timeout occurs, the value of TimeoutInterval is doubled to avoid another timeout for the same packet. This provides some congestion control as well - the more timeouts occur, the longer TCP will wait before retransmission. Once the troublesome packet goes through, TimeoutInterval is restored to its last computed value. EstimatedRTT is not computed for packets which are retransmitted, so this doubling does not affect the value.

Informally at least, delay and network latency are the same thing: the time that it takes a packet to get from point A to point B.For lots of reasons, that's not a single number, that's really a probability distribution function, with a lower bound given by the laws of physics.Ideally, you'd have a pure spike at this lower bound, but practical reasons turn that into some other kind of curve. As in statistics, the variance is simply telling you about how 'wide' or 'broad' that curve is, and is trying to capture how much change there is in the network latency.