Counting citations per paper across the whole graph has to touch all 352,807 edges [2]. Divide the two slow aggregation times by that edge count and you get about 11 microseconds per edge for ArangoDB Oasis and about 5 for CognoDB [1][2]. On a graph this size, per-edge costs in the microseconds are what turn a query into a batch job, which is roughly how the author frames the difference between a full scan and a point lookup [16]. He checked the boring explanation first: db.explain() showed the edge index in use with no planner warnings [11].
His reading of the concurrency result is a protocol ceiling. ArangoDB Oasis is the only platform in the set speaking HTTP/REST, where three others use Bolt and FalkorDB uses native RESP [12], and he says plainly that he cannot prove root cause from a free-tier black box [13]. I would take it on those terms. The evidence has the right shape for it, because per-request overhead is invisible in one timed query and expensive when forty clients are pushing at once [8].
The flatline itself rests on a thin sample. Two ten-second runs at those rates come to roughly 310 ArangoDB operations in total [4]. The reported 1.05x does not quite agree with the printed integers either, whose ratio is 1.07 [5], so there were decimals behind them. The direction is not in question; a 4x load increase observed over twenty seconds of wall clock is still a smoke test rather than a capacity model.
For any of this to transfer, two things have to hold. Your traffic has to be dominated by the shape that decided the ranking, and the gap between the fast pair and the slower three on hops and lookups is roughly 1.6x to 2.2x at the reported boundaries [7], which stops being a procurement criterion the moment your latency budget is 250ms. Your graph also has to look like this one, averaging 12.7 edges per node [3], uniform enough that a three-hop traversal is a bounded amount of work rather than an accidental scan through a supernode.
The finding I would act on is the cheapest one to verify. CognoDB speaks Bolt and Cypher, and the author repointed the same client code at it by changing a URI, a username and a password [15]. That is what makes this argument operational: if a candidate costs three environment variables to try, you can run the suite against your own query mix instead of adjudicating someone else's. Worth noting that the author was asked to benchmark CognoDB against the other four [1], and it landed mid-pack on traversals and lookups [14] with the second-slowest aggregation of the five [8]. The post does not say who asked, and the result reads like measurement rather than promotion.
