Feature request: an option to skip per-method docstring generation in build()
Problem
build() (and lazy sub-resource construction) generates a fully-expanded, recursively
pretty-printed prototype of each method's response schema and attaches it as
method.__doc__ (discovery.createMethod → schema.Schemas.prettyPrintSchema /
prettyPrintByName). For APIs with large, deeply-nested schemas this is very expensive, and it
is paid every time a service/resource is constructed.
Concrete numbers from profiling Sheets v4 (google-api-python-client==2.198.0, Python 3.13),
measured with RSS (no tracemalloc, to avoid its overhead):
- Building the service, then touching one sub-resource (
service.spreadsheets(), no API
call): ~66 MB.
- Of that, ~99.9% is the docstring schema expansion — no-oping
prettyPrintSchema/
prettyPrintByName drops it to ~1 MB. The .spreadsheets() methods themselves are ~24 KB.
- The docstrings are only useful for interactive
help(); in a server they are never read.
Impact
In a concurrent server (a fresh service built per request, common with per-user credentials),
these allocations are not shared across in-flight requests. 8 concurrent Sheets requests
each build ~66 MB of docstrings simultaneously ≈ 530 MB peak, which OOM-kills a
memory-limited container. This is the concurrent-peak sibling of the long-standing
reference-cycle memory issue in #535 (whose recommended fix — build/reuse a single service — is
not always feasible when credentials differ per request).
Request
A supported way to skip docstring generation at build time, e.g.:
build("sheets", "v4", credentials=creds, generate_docstrings=False)
# or a module/env toggleToday the only options are to monkeypatch Schemas.prettyPrintSchema/prettyPrintByName
(fragile across versions) or fork. A first-class flag would let memory-constrained / high-
concurrency deployments opt out of documentation strings they never use.
Environment
google-api-python-client==2.198.0, Python 3.13- Reproly: build any large-schema API (Sheets v4), touch a sub-resource, measure RSS; repeat
concurrently to see the multiplier.
Related: #535 (memory from repeated build() / reference cycles).
Feature request: an option to skip per-method docstring generation in
build()Problem
build()(and lazy sub-resource construction) generates a fully-expanded, recursivelypretty-printed prototype of each method's response schema and attaches it as
method.__doc__(discovery.createMethod→schema.Schemas.prettyPrintSchema/prettyPrintByName). For APIs with large, deeply-nested schemas this is very expensive, and itis paid every time a service/resource is constructed.
Concrete numbers from profiling Sheets v4 (
google-api-python-client==2.198.0, Python 3.13),measured with RSS (no tracemalloc, to avoid its overhead):
service.spreadsheets(), no APIcall): ~66 MB.
prettyPrintSchema/prettyPrintByNamedrops it to ~1 MB. The.spreadsheets()methods themselves are ~24 KB.help(); in a server they are never read.Impact
In a concurrent server (a fresh service built per request, common with per-user credentials),
these allocations are not shared across in-flight requests. 8 concurrent Sheets requests
each build ~66 MB of docstrings simultaneously ≈ 530 MB peak, which OOM-kills a
memory-limited container. This is the concurrent-peak sibling of the long-standing
reference-cycle memory issue in #535 (whose recommended fix — build/reuse a single service — is
not always feasible when credentials differ per request).
Request
A supported way to skip docstring generation at build time, e.g.:
Today the only options are to monkeypatch
Schemas.prettyPrintSchema/prettyPrintByName(fragile across versions) or fork. A first-class flag would let memory-constrained / high-
concurrency deployments opt out of documentation strings they never use.
Environment
google-api-python-client==2.198.0, Python 3.13concurrently to see the multiplier.
Related: #535 (memory from repeated
build()/ reference cycles).