ObliviOS Developer portal Project site

Home / Decisions / ADR-164

The whole-branch review of the server places (ADR-161) corrects four of its statements and replaces its rule for networks that steer by DNS.

Accepted, recorded 13 October 2026.

Decision
ADR-164
Date
2026-10-13
Status
AcceptedAccepted (owner, 2026-10-13; the owner left the list of (3) and the reading of (7) to the assistant); the list is read against the real inputs before a file is shipped
Mentions
ADR-161
Mentioned by
ADR-161

1Decision

The whole-branch review of the server places (ADR-161) corrects four of its statements and replaces its rule for networks that steer by DNS. (1) What is kept for a service is the kind, country, city, coordinates rounded to 0.01 degree, AS number and operator name of the server, never the address; this replaces "only country and ASN are kept" in ADR-161 (2).

(2) The lookup runs on the service thread of the Data Map service, only for a service that has no place to draw yet; the callback of netd only hands an address on; this replaces "in the collector at the event" in ADR-161 (2).

(3) A network is treated as steered by DNS when its operator name holds, as whole words in any case, one of amazon, akamai, microsoft, google, cloudflare, fastly, facebook, meta platforms, apple, edgecast, edgio, limelight, cdn77, stackpath, imperva, incapsula; this replaces the list "Amazon, Akamai, Microsoft" of ADR-161 (3), and tools/audit-steered.py prints the matches so that the list is read against the inputs before a file is shipped.

(4) A server place is never merged with another and is named after its own city, or its country when the data has no city; a service has one answer, the best of what its entries said (a place that can be drawn over one that cannot, the first at equal rank). (5) A city is drawn when the city data and the country data name the same country or the country data names none; a country alone is drawn at its anchor; where they name two countries there is no place.

(6) The choice to keep server places in the history is a file of its own, so that an earlier build reads the settings file as before (it does not save the history: an earlier build reads no row of seventeen fields and starts the history again). The choice belongs to the history: it cannot be switched on with the history off and goes off with it. (7) The attribution to DB-IP is plain text that names db-ip.com, because the app opens no link (MAP-011).

(8) A reply to the app is bounded in bytes (entries 512 KB, 16 pieces of evidence for each of at most 300 links) and the history file is written under 56 MB by leaving out the oldest days; a place that can be drawn also wins when the days are put together. (9) The address file is mapped read-only; a damaged page of /system is not treated apart from the other mapped system files.

2Why and evidence

The review found that a card named one city for servers in three, that the sheet listed companies for a server place, that one long row made the history unreadable, that the threat model claimed two sources where one can draw a place, and that the steered list missed the operators that the measurement names (Google 49 of 93, Microsoft 14 of 16 outside the census; the two country sources agree for 70% of the large steered networks, against 93% elsewhere).

3Consequences

The image grows by about 48 MB (country and ASN for IPv4 and IPv6, the city part for IPv4 [measured]), not the 8 MB of ADR-161. Fewer addresses get a place: services of Google, Cloudflare and the other added operators show "behind <operator>". The file of the first build (47 942 775 bytes) was made with the earlier rule and inputs that are not archived, and is rebuilt before it is landed; its figures (82 172 addresses compared) belong to that build. The list of (3) is chosen from names and [assumption] until it is read against the real inputs.

4Review when

The audit of the real inputs shows a false match or a missing edge operator; the owner rejects the reading of (7).

Read from docs/governance/decision-log.md, entry ADR-164.