Consider "bazel sync" instead of "bazel fetch"

[coryan]

We use bazel fetch in our (Bazel) CI builds to prevent flakes caused by failed downloads. It seems like we could simplify this by using bazel sync instead. fetch requires listing the rules that will be executed, and some indirect dependencies. sync seems to download things we do not need, e.g., the SDKs for Windows when compiling on Linux.

[devjgm]

I think we want to keep bazel fetch not bazel sync. The later downloads more deps than are necessary. However, I think we can simplify our bazel fetch invocation from

google-cloud-cpp/ci/cloudbuild/builds/lib/bazel.sh

Lines 42 to 47 in a6329c1

	"ci/retry-command.sh" 3 120 bazel fetch ... \
	@local_config_platform//... \
	@local_config_cc_toolchains//... \
	@local_config_sh//... \
	@go_sdk//... \
	@remotejdk11_linux//:jdk

To simply bazel fetch ....

I think the simpler invocation is supposed to download all the deps that are necessary. And in an experiment on my local machine, it appeared to do just that. I did a clean bazel build ..., and looked at all the files downloaded in ~/.cache/bazel/_bazel_jgm/cache/repos/. Then I expunged everything and did a bazel fetch ..., and it fetched everything that the bazel build fetched.

[devjgm]

Well, I just found #5346, which seems to suggest that I'm wrong... let me do a bit more investigation.

[devjgm]

I think we should be able to only specify bazel fetch ... and have it get everything, but the PR above is correct that we need to list more. I think that's a bazel bug and filed: bazelbuild/bazel#13847.

Regardless, for the purpose of this issue, we want to keep bazel fetch. bazel sync is not what we want because it fetches way too much.