[C++] S3FileSystem construction causes IDMS lookups for region even when specified

[apmorton]

Describe the bug, including details regarding any error messages, version, and platform.

When constructing a S3FileSystem object with a region explicitly specified arrow will still end up causing aws metadata lookup operations, unless explicitly (and globally) disabled with the AWS_EC2_METADATA_DISABLED environment variable.

This goes against the documented (and I believe intended?) behavior of the region kwarg:

AWS region to connect to. If not set, the AWS SDK will attempt to determine the region using heuristics such as environment variables, configuration profile, EC2 metadata, or default to ‘us-east-1’ when SDK version <1.8.

pyarrow.fs.S3FileSystem(
    access_key='key',
    secret_key='secret',
    endpoint_override='https://my.appliance.uri',
    region='region',
)

using py-spy I can observe the following stack:

    Aws::Http::CurlHttpClient::MakeRequest (libaws-cpp-sdk-core.so)
    Aws::Internal::AWSHttpResourceClient::GetResourceWithAWSWebServiceResult[abi:cxx11] (libaws-cpp-sdk-core.so)
    Aws::Internal::EC2MetadataClient::GetCurrentRegion[abi:cxx11] (libaws-cpp-sdk-core.so)
    Aws::Client::ClientConfiguration::ClientConfiguration (libaws-cpp-sdk-core.so)
    Aws::S3::S3ClientConfiguration::S3ClientConfiguration (libaws-cpp-sdk-s3.so)
    __gnu_cxx::new_allocator<arrow::fs::S3FileSystem::Impl>::construct<arrow::fs::S3FileSystem::Impl, arrow::fs::S3Options const&, arrow::io::IOContext const&> (libarrow.so.1801.0.0)
    arrow::fs::S3FileSystem::S3FileSystem (libarrow.so.1801.0.0)
    arrow::fs::S3FileSystem::Make (libarrow.so.1801.0.0)
    S3FileSystem___init__ (pyarrow/_s3fs.cpython-311-x86_64-linux-gnu.so)

This is caused by default construction of S3ClientConfiguration in ClientBuilder.
On our machines (which aren't in aws and have no idms running) this takes 6+ seconds.

A workaround is something as follows:

#ifdef ARROW_S3_HAS_S3CLIENT_CONFIGURATION
  Aws::S3::S3ClientConfiguration client_config_{Aws::Client::ClientConfigurationInitValues{false}};
#else
  Aws::Client::ClientConfiguration client_config_{Aws::Client::ClientConfigurationInitValues{false}};
#endif

which disables idms during configuration construction.

Some additional work would be required to add back in IDMS lookup of region when otherwise not specified.

Component(s)

C++

[pitrou]

Sorry for the delay, I've only just noticed this issue. I will take a look, but this sounds like a reasonable suggestion to me.

[pitrou]

Also, what do you think of using the "standard" defaults mode instead of the "legacy" defaults mode as outlined here:
https://docs.aws.amazon.com/sdkref/latest/guide/feature-smart-config-defaults.html ?

[pitrou]

@apmorton I have tried both your solution and another similar solution, and I don't see any obvious timing or behavior change on my local machine. Am I missing something?

Here are the two diffs I'm trying:

diff --git a/cpp/src/arrow/filesystem/s3fs.cc b/cpp/src/arrow/filesystem/s3fs.cc
index 62daab249c..919fdb8908 100644
--- a/cpp/src/arrow/filesystem/s3fs.cc
+++ b/cpp/src/arrow/filesystem/s3fs.cc
@@ -1069,7 +1069,13 @@ class EndpointProviderCache {
 
 class ClientBuilder {
  public:
-  explicit ClientBuilder(S3Options options) : options_(std::move(options)) {}
+  ClientBuilder() = delete;
+
+  explicit ClientBuilder(S3Options options)
+      : options_(std::move(options)),
+        // The ClientConfiguration constructor always does a current region lookup
+        // via EC2 metadata, unless IMDS is disabled (GH-46214).
+        client_config_(Aws::Client::ClientConfigurationInitValues{false}) {}
 
   const Aws::Client::ClientConfiguration& config() const { return client_config_; }
 

diff --git a/cpp/src/arrow/filesystem/s3fs.cc b/cpp/src/arrow/filesystem/s3fs.cc
index 62daab249c..65918c0025 100644
--- a/cpp/src/arrow/filesystem/s3fs.cc
+++ b/cpp/src/arrow/filesystem/s3fs.cc
@@ -1069,7 +1069,14 @@ class EndpointProviderCache {
 
 class ClientBuilder {
  public:
-  explicit ClientBuilder(S3Options options) : options_(std::move(options)) {}
+  ClientBuilder() = delete;
+
+  explicit ClientBuilder(S3Options options)
+      : options_(std::move(options)),
+        // The ClientConfiguration constructor always does a current region lookup
+        // via EC2 metadata, unless IMDS is disabled (GH-46214).
+        client_config_(/*useSmartDefaults=*/true, "legacy",
+                       /*shouldDisableIMDS=*/true) {}
 
   const Aws::Client::ClientConfiguration& config() const { return client_config_; }
 

[pitrou]

Ok, ignore the previous comment. I might have had a stale shared library somewhere that made my changes being ignored.