State of the art, per sub-question

Each sub-question has two records, kept apart: the most adopted practice (what people use) and the best evidence (what a study measured). Upper bounds stay separate from lower bounds; the first says what has been achieved, the second what is impossible.

Sub-question Most adopted Best evidence Measured how
Q1 Principles of system design Lampson 1983 hints; end-to-end argument (Saltzer 1984) none found that tests a principle argument and experience only
Q2 Consistency and availability binary CP or AP framing (Primer 2026) Brewer 2012 and Kleppmann 2015 (arguments on the framing); Bailis 2012 (staleness) argument; model fitted to production latencies of two firms
Q3 What designs get wrong postmortem practice (Beyer 2016); Jepsen 2026 Yuan 2014; Alquraan 2018; Gunawi 2016 198 and 136 failures, read and partly reproduced; 597 outages from public reports
Q4 Capacity and load estimation back-of-the-envelope in primers (Primer 2026) Dean 2013 (fan-out arithmetic and one measured table); no study of estimate accuracy arithmetic; one Google service
Q5 Interfaces between services company rule sets (Google AIP 2025; Zalando 2026; Microsoft API 2026) Stylos 2007 (30 programmers, user study); Piccioni 2013 (25); Murphy 2018; Kotstein 2021; Rauf 2019 user studies on library APIs; 24 interviews; 8-expert Delphi; mapping of 47 studies
Q6 Checking a design before building design review; TLA+ at AWS (Newcombe 2015); simulation (Zhou 2021) Newcombe 2015 and Bornholt 2021 (owners' reports, with counts); Bogli 2024 (review) experience, no control; 16 industrial papers
Q7 Systems engineering standards ISO 2023b (15288); INCOSE 2023 (not opened) Elm 2012 survey with statistical links
Q8 Interview books and primers Primer 2026: 376k stars none: no study compares them with research counts of stars only
Models as implementer none Cheng 2026 (benchmark, no results in abstract); Sahu 2026 benchmark design; argument with a tool

Lower bounds, upper bounds

  • Lower bound: no web service can have consistency, availability and partition tolerance together, proved in the asynchronous network model (Gilbert 2002, abstract); Kleppmann 2015 shows the result depends on its definitions.
  • Lower bound, practical: latency is traded against consistency even without a partition (Abadi 2012, PACELC).
  • Best case in two modelled profiles: in Dynamo-style quorums fitted to LinkedIn latencies (N=3, R=W=1), 99.9% of reads were predicted consistent after 1.85 ms to 45.5 ms (Bailis 2012). This is one configuration, not a bound.
  • Upper bound: 93% of partition failures studied could be reproduced by fault injection (Alquraan 2018, Finding 13).

Adoption counts, descriptive, taken 2026-10-29

Count Value Caveat
Stars of the system design primer 375,957 created 2017-02-26; last push 2026-09-15
Forks of the primer 59k page figure, rounded
Stars of the Microsoft REST guidelines 23.3k page figure, rounded
Jepsen analyses listed 51 entries for 39 systems 2013-05-18 to 2026-10-20; counted by the page reader
Azure cloud design patterns 44 counted in the catalog table
Stars of another primer (ByteByteGo 2023) 90,808 repository count only; not read

Reading the table

  • Most adopted and best evidence are different holders on every row where both exist.
  • Where the research record is empty (Q1, Q4, Q8), this search found nothing; that is not proof nothing exists.