fix(ingest): loosen confluent-kafka dep requirement

metadata-ingestion/setup.py

@@ -64,12 +64,40 @@ def get_long_description():
 }
 
 kafka_common = {
+    # The confluent_kafka package provides a number of pre-built wheels for
+    # various platforms and architectures. However, it does not provide wheels
+    # for arm64 (including M1 Macs) or aarch64 (Docker's linux/arm64). This has
+    # remained an open issue on the confluent_kafka project for a year:
+    #   - https://github.com/confluentinc/confluent-kafka-python/issues/1182
+    #   - https://github.com/confluentinc/confluent-kafka-python/pull/1161
+    #
+    # When a wheel is not available, we must build from source instead.
+    # Building from source requires librdkafka to be installed.
+    # Most platforms have an easy way to install librdkafka:
+    #   - MacOS: `brew install librdkafka` gives latest, which is 1.9.x or newer.
+    #   - Debian: `apt install librdkafka` gives 1.6.0 (https://packages.debian.org/bullseye/librdkafka-dev).
+    #   - Ubuntu: `apt install librdkafka` gives 1.8.0 (https://launchpad.net/ubuntu/+source/librdkafka).
+    #
+    # Moreover, confluent_kafka 1.9.0 introduced a hard compatibility break, and
+    # requires librdkafka >=1.9.0. As such, installing confluent_kafka 1.9.x on
+    # most arm64 Linux machines will fail, since it will build from source but then
+    # fail because librdkafka is too old. Hence, we have added an extra requirement
+    # that requires confluent_kafka<1.9.0 on non-MacOS arm64/aarch64 machines, which
+    # should ideally allow the builds to succeed in default conditions. We still
+    # want to allow confluent_kafka >= 1.9.0 for M1 Macs, which is why we can't
+    # broadly restrict confluent_kafka to <1.9.0.
+    #
+    # Note that this is somewhat of a hack, since we don't actually require the
+    # older version of confluent_kafka on those machines. Additionally, we will
+    # need monitor the Debian/Ubuntu PPAs and modify this rule if they start to
+    # support librdkafka >= 1.9.0.
+    "confluent_kafka>=1.5.0",
+    'confluent_kafka<1.9.0; platform_system != "Darwin" and (platform_machine == "aarch64" or platform_machine == "arm64")',
     # We currently require both Avro libraries. The codegen uses avro-python3 (above)
     # schema parsers at runtime for generating and reading JSON into Python objects.
     # At the same time, we use Kafka's AvroSerializer, which internally relies on
     # fastavro for serialization. We do not use confluent_kafka[avro], since it
     # is incompatible with its own dep on avro-python3.
-    "confluent_kafka>=1.5.0,<1.9.0",
     "fastavro>=1.2.0",
 }