I used Claude to investigate some details on database errors, and gathered some information for review:

SYMPTOM

Our MySQL connections intermittently stall for ~45 seconds. The database is

effectively idle when this happens. On 2026-07-13 (UTC) our app logged two

episodes via HikariCP leak detection (threshold 30s):

02:18:36 -> 02:19:21 ~45s three separate connections, ALL released in the

                            same millisecond (02:19:21.103)

13:35:03 -> 13:35:48 ~45s single connection

Separately, a single-row INSERT into a table containing one row, run from a

client via the TCP proxy (gondola.proxy.rlwy.net:38846), took ~10 seconds.

WHAT WE HAVE ALREADY RULED OUT

This is not load, resources, or a slow query:

• MySQL CPU: 0 vCPU (0%) of 32 vCPU
• MySQL memory: 707 MB of 32 GB
• Volume: 1.29 GB of 48.8 GB
• App traffic: ~3 HTTP requests/hour at the time
• Innodb_log_waits = 0 (redo log not a bottleneck)
• Innodb_data_pending_fsyncs = 0 (storage is NOT stalling)
• Threads_connected = 31 (Max_used_connections 62 of max 151)
• Aborted_clients = 26 (clients dying without a clean close)

Config is the stock Railway MySQL template, untuned.

WHY WE THINK IT IS THE NETWORK / HOST, NOT MYSQL

• Zero pending fsyncs and zero log waits, at 0% CPU, rules out disk and CPU.

• Both connection paths are affected: the private network

  (mysql.railway.internal) AND the public TCP proxy. Those are different

  routes; the common factor is the MySQL instance's host/network.

• Three connections unblocking in the SAME MILLISECOND is characteristic of a

  connection-path stall clearing, not of a database working through a queue.

• Aborted_clients climbing is consistent with sockets being dropped.

IMPACT

Certain Inserts take much longer than expected.

WHAT WE ARE ASKING

1. Were there host events — migration, throttling, noisy neighbour, network

   maintenance — affecting this MySQL instance during those two windows?

2. Are there known private-network / TCP-proxy issues in US West around

   2026-07-13?

3. Is there anything on your side we can monitor or alarm on so we see this

   before a customer does?

Happy to provide raw logs or run any diagnostics you need.