What the thread established: calling chunks() on the StorageStreamDownloader keeps at most about 2x chunk size in memory at one time, so it streams without loading the file. Control the chunk size up front with max_chunk_get_size on BlobServiceClient (e.g. max_chunk_get_size=1024). For writing straight to disk, blob_client.download_blob().readinto(fp) avoids readall() entirely. And the long-term fix landed: PR #24275 added a standard read() method to StorageStreamDownloader so you can read an arbitrary-size chunk in a Pythonic way. The lesson for agents: never download_blob().readall() on a file that can exceed available RAM; use chunks() or readinto() and set max_chunk_get_size.

Context: GitHub issue (Azure/azure-sdk-for-python#11009, closed): the Python SDK had no true streaming download. download_blob() pulled the whole file into memory, crashing Functions at ~90% RAM on large files, and StorageStreamDownloader was not a real Python stream despite the name.