Repository navigation
Consistent style in the ABI (expanded vs minimal) #80
Description
During the initial development, we went through a lot of back-and-forth both expanding and minimizing the ABI surface (e.g. get_buffer_bytes(buffer_type) vs get_http_request_body_chunk() and similar, set_shared_data(NULL) vs remove_shared_data()), both internally and during Envoy upstreaming process.
The version that was initially merged into Envoy, and which the ecosystem ossified around, ended up with some unfortunate inconsistencies (using both: expanded proxy_{get,add,replace,remove}_header_map_value, and minimal proxy_set_buffer_bytes(start, offset) or proxy_set_shared_data(NULL)).
We should clean this up and stick to one style.
While I don't have strong preference one way or the other, note that splitting features into groups (#57) and having specialized context types with their own hostcalls (#9) probably means that we should be expanding the ABI surface instead of minimizing it.
@mpwarres @martijneken @leonm1 @wbpcode @thibaultcha any thoughts?
Activity
mpwarres commentedon May 29, 2025
I don't feel strongly, but as discussed offline, my preference would be for separate hostcalls, to make usage/dependency visible to static analysis, and also to leave open flexibility for variation if needed.
PiotrSikora commentedon Mar 19, 2026
Per discussion from the ABI sync meeting on 2/19, we've decided to go with the "minimal" variants, otherwise the number of hostcalls would increase significantly.
For example, expanded
proxy_get_header_map_pairs(proxy_map_type_t)would require 8x hostcalls -proxy_get_request_header_map_pairs,proxy_get_response_header_map_pairs, ..., expandedproxy_get_buffer_bytes(proxy_buffer_type_t)would require 9x hostcalls, and so on... Same for setters.Simiarly, using
0(NULL) in place of return buffer in getters will be used forexistscalls, and using0(NULL) in place of new value in setters will be used fordeletecalls.