Log-linear is a specific form of HDR. Adaptive sampling still needs somewhere to put the data, so that's not a "representation." The choices are: 1) HDR, 2) T-digest, 3) moment sketches, 4) raw samples. 1/https://twitter.com/pmoust/status/1124996717777174529 …
-
Prikaži ovu nit
-
The idea of "not doing it at all" is valid, it just means a vendor doesn't want to provide an answer to your questions. However, anything framed around quality-of-service, SLI/SLO/SLA, or availability requires questions like this. So, avoid at your own risk. 2/
1 reply 1 proslijeđeni tweet 6 korisnika označava da im se sviđaPrikaži ovu nit -
#4 is expensive, you can sample, but it is still expensive. The sampling is better for exemplar selection than for actual math. This leaves of us with 1, 2, or 3. (and, again, log-linear is a form of 1). 3/
1 reply 0 proslijeđenih tweetova 1 korisnik označava da mu se sviđaPrikaži ovu nit -
My favorite from a sheer elegance PoV is #3 as moment sketches are beautiful. Both #1 and #2 are similar in that they trade accuracy over parts of the cumulative density function (CDF) they are approximating. This introduces error that is hard to reason about. 4/
0 proslijeđenih tweetova 3 korisnika označavaju da im se sviđaPrikaži ovu nit -
(IMHO) 2 and 3 are "harder to comprehend" so if you expose the histogram to end users, it is harder to explain how what you show represents the reality of the underlying data. HDR, on the other hand, is dead-simple to understand. 5/
1 reply 0 proslijeđenih tweetova 1 korisnik označava da mu se sviđaPrikaži ovu nit -
When the HDR technique is universally consistent within an application (e.g. Circonus log-linear histograms), it affords mergeability guarantees that allow for the accumulation of arbitrary histograms without new error or changing of the user's mental model. 6/
1 reply 0 proslijeđenih tweetova 2 korisnika označavaju da im se sviđaPrikaži ovu nit -
Summary, if you're modeling for point-in-time (now) decision making, #2 and #3 are exceptionally good. If you are presenting to a human where they drive the time frames arbitrarily (monitoring/reporting/planning), then #1 is best. #4 is just too expensive. 7/
1 reply 0 proslijeđenih tweetova 3 korisnika označavaju da im se sviđaPrikaži ovu nit
The @Circonus circllhist implementation constrains the representation such that all histograms everywhere are mergeable in a way that reduces surprise to the end consumer of that data (devs, operators, data scientists, etc.). This makes it more suitable for general use.
END/
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.