[dkropachev]

Summary

Sync clients and resources could open a fresh TCP connection on every request when cycling through more than ten nodes because botocore leaves urllib3's host-pool cache at its default size. Grow the existing cache to cover seed, routing, and key-affinity nodes, including topology changes and lazily created proxy managers.

Keep max_pool_connections as the per-host connection limit and preserve the existing transport's TLS, proxy, retry, and timeout settings. Regression coverage checks that two passes through eleven HTTP endpoints reuse eleven TCP connections.

Fixes #159. Synchronous discovery pooling is tracked separately in #160.

[qodo-scylladb]

Code Review by Qodo

---

Resolved findings

1. Topology changes can evict live pools ✓ Resolved

Description

grow_pools sizes the cache from current node snapshots but does not count the destination already
selected for the request being sent. If discovery replaces an eleven-node snapshot while a request
using its earlier query plan is in flight, that old destination can consume one of the eleven slots
needed by current nodes and evict a reusable pool.

Code

alternator/_sync_sdk.py[R36-39]

+    def grow_pools(**kwargs: object) -> None:
+        hosts = set(config.seed_hosts)
+        hosts.update(manager.nodes.nodes)
+        hosts.update(manager.affinity_nodes.nodes)

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The before-send cache size excludes a destination selected from a previous topology snapshot. An in-flight request to that destination can evict a pool belonging to the current snapshot.

## Fix Focus Areas
- alternator/_sync_sdk.py[36-51]
- tests/unit/test_sync_connection_pools.py[26-64]

## Recommended Fix
Include the before-send request’s destination in the host set used for sizing. Add a test that changes topology after routing selects a host but before that request is sent.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

2. Concurrent proxy requests can fail ✓ Resolved

Description

grow_pools iterates _proxy_managers.values() without synchronizing with lazy proxy-manager
creation. When concurrent first requests use a proxy, insertion during that iteration can raise
RuntimeError; a manager created after the snapshot can also serve a burst with the default
ten-host cache.

Code

alternator/_sync_sdk.py[42]

+        pool_managers = [transport._manager, *transport._proxy_managers.values()]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Proxy managers are created lazily after the before-send callback, while the callback iterates their registry without synchronization. Concurrent initial requests can race with that iteration or use an unsized manager.

## Fix Focus Areas
- alternator/_sync_sdk.py[40-51]
- tests/unit/test_sync_connection_pools.py[26-64]

## Recommended Fix
Arrange for each proxy manager to be sized when it is created, with access to the manager registry synchronized against sizing. Add coverage for concurrent initial proxy requests rather than only a proxy manager created before the event.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

---

Tip of the day

💡 Did you know, you can describe a rule in plain language on the Rules page and Qodo drafts it for you

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

---

Powered by Qodo

[dkropachev]

In reply to

Include the before-send request’s destination in the host set used for sizing.

Cache misses now account for the selected destination and protect pools for the seed and observed routing/affinity snapshots. Snapshot reads, recency promotion, and insertion share the cache lock. Regressions cover topology replacement after routing, waiting lookups, and repeated retired destinations with direct and HTTPS proxy managers.

Arrange for each proxy manager to be sized when it is created, with access to the manager registry synchronized against sizing.

The original SDK proxy getter now serializes lazy construction and installs the sizing hook before the manager’s first pool lookup, without iterating the registry. Added concurrent first-request coverage for clients and resources, plus a regression preserving SDK proxy-error mapping and standard retries. The per-host connection limit is unchanged.