Operations
Known issues
An honest, operational view: what is rough, what is preview or rolling out, and what has a workaround in the current 5.0.x line.
Current limitations
| Area | What to know |
|---|---|
| Requests forwarded without the user's credential | A request that a cluster node forwards on a user's behalf without the user's credential can still be treated as node traffic on a few paths (chat that crosses nodes from inside the system). On the known paths this affects attribution to a user and tenant, not permissions. A fix is planned. See the 5.0.177 release notes. |
| Memory: bulk deletion through the chat (5.0.173, 5.0.174) | The chat assistant's forget tool can clear a whole memory namespace for an ordinary user, including other users' private memories. The 5.0.172 fix for this is missing in 5.0.173 and 5.0.174. The direct memory API is not affected. Fixed in 5.0.175; update to it. See the release notes. |
| Rolling out an update across a cluster | Update the edge nodes and the controllers first, then the other nodes (from 5.0.177, forwarding nodes pass on who the caller is, and an older forwarding node in front of updated nodes loses that). While nodes run different versions, a knowledge-base search that reaches an older data node can miss that node's results; they are reported as an error. Update all nodes of a cluster to the same release. |
| Standby data node: knowledge bases (5.0.177) | A standby data node does not receive the knowledge-base content that existed before replication was switched on (or that was written while replication was off), and a standby added later gets none of it either. New writes do reach it. The replication status and the check before a promotion still report such a standby as caught up, so do not promote it, manually or automatically, if you need your knowledge bases there. Memory and tenant files are not affected. Model copying can also stall the other copy rounds for hours and fill the standby's disk; do not start a manual file sync while one is running. In 5.0.177 model copying cannot be switched off through the settings, so if this affects you, switch replication off (mode none) until the fix. A fix is in progress. |
| Data-node failover | Automatic failover to the standby data node is off by default; turn it on only after testing on your own nodes. A failed primary that comes back is fenced, but tenant files written directly into its storage while fenced are not refused, only not shipped further. |
| Local models | The CPU package runs local models only on CPUs with the needed instructions: AVX2 on x86_64, and on aarch64 those of the Raspberry Pi 5 and newer. On other CPUs, such as a Raspberry Pi 4, those requests answer 503 with the reason, and the rest of the node keeps running. |
| Logs: NetFlow and limits | NetFlow and IPFIX volumes are kept as reported by the exporter, not exact. Sign-in and password-reset rate limits count per node in memory, so a restart resets the counts. |
| Knowing which version you are actually on | The CPU package (eldric-aios) and the GPU package (eldric-aios-cuda) are published independently and can sit at different versions, so the “Applies to” note above may not match your node. Check with rpm -q eldric-aios eldric-aios-cuda, and see release notes for why the two channels drift and what dnf update does on each host. Note that eldric --version does not report it on Linux: there /usr/bin/eldric is a symlink to the daemon, which does not accept the flag and exits with usage. |
| High-concurrency knowledge-base search | Tune namespace size and monitor latency under load. |
| Advanced retrieval & memory | Some controls are opt-in; enable per tenant where marked preview. |
| Role-based access control | Enforcement is opt-in and off by default — an installation restricts nobody until an administrator enables it and switches the mode. Turn it on deliberately, and run the audit mode first: see Turning on role enforcement. |
| Packaging | Supported on Fedora 43 and Fedora 44, on x86_64 and aarch64; the GPU package is built for Fedora 43 on x86_64 only. Fedora 42 and RHEL-family systems are not supported at the moment. If you installed Eldric on Fedora 42 or a RHEL-family system, or on Fedora 44 before 5 October 2026, you are probably running an old version: check with rpm -q eldric-aios. From 5.0.175 the full package is built for aarch64 too; an older ARM64 minimal package (version 5.0.0) is still listed in the repository and is not updated, so do not use it. Use the packages on repo.eldric.ai. |
Preview & rolling out
Enable preview features deliberately per tenant or deployment, and describe them as preview in customer handoff — do not present them as generally available:
- World-model / scientific rendering — a scoped visualization of named domains through a known model (not prediction); demonstrable today, hardening in the 5.0.x line.
- Smart-memory inference — only in the separate native inference daemon that comes with the GPU package; an administrator switches it on per daemon (
--enable-mhc-retrieval). Off by default. - Mobile apps — the iOS app is in internal testing and not offered to customers yet. There is no Android app on offer.
- DevToolbox — available in the macOS app (Tools menu). In the iOS app it is a preview: the prompt playground, agent debugger and performance profiler do not run yet.