Proposal: When a client is pushed back onto the pool, reduce the size of large connection buffers.

Background & Motivation

HELP-78501 reports observed memory growth in a long-lived process. This appears to be due to reading large responses on a buffered stream. The long-lived buffered stream does not later reduce the size until the pool is destroyed.

An example creating many cursors reading large responses reproduces the issue. Here is a simplified snippet:

mongoc_client_t *client[100];
for (int i = 0; i < 100; i++) {
    client[i] = mongoc_client_pool_pop (pool);
}

for (int i = 0; i < 100; i++) {
    // Server replies may be 16MB each.
    find_all_documents (client[i], "db", "big_collection");
}

// 1600 MB allocated and will not be freed until `mongoc_client_pool_destroy`.

macOS Instruments shows many 16MB buffers allocated after calls to find_all_documents complete:

After calling mongoc_client_pool_destroy, memory usage lowers to near 0.

The default stream initiator creates buffered streams.