The BaseDataVariableState class consumes too much memory

[kimdiego2098]

Version: 2.0.173-preview

All objects in the snapshot
Type, Objects, Bytes, Minimum retained bytes
ThingsGatewayRuntime.Plugin.OpcUa.OpcUaTag, 100014, 91212768, 113615904

internal sealed class OpcUaTag : BaseDataVariableState
{
    internal OpcUaTag(NodeState parent) : base(parent)
    {
    }

    /// <summary>
    /// 叶子变量不包含子节点，直接更新自身状态以避免基础实现创建空子节点列表。
    /// </summary>
    /// <param name="context">当前 OPC UA 系统上下文。</param>
    /// <param name="statusCode">需要设置的状态码。</param>
    /// <param name="timestamp">状态对应的源时间戳。</param>
    public override void SetStatusCode(
        ISystemContext context,
        StatusCode statusCode,
        DateTimeUtc timestamp)
    {
        StatusCode = statusCode;
        if (timestamp != DateTimeUtc.MinValue)
        {
            Timestamp = timestamp;
        }
    }

    internal long Id { get; set; }

    internal bool IsDataTypeInit { get; set; }
}

The BaseDataVariableState class consumes too much memory. When the software needs to provide hundreds of thousands of OPC variables, the memory occupied is quite exaggerated

[romanett]

@kimdiego2098 Do you have a proposal what to change?

Copilot has the following, but we would need to discuss if we want to change those:

Plausible mitigations

• Lazily allocate the locks (or consolidate into a single lock / lock on the collection object) — biggest structural win, removes ~6 objects/node.

• Move the rarely-used delegate handlers off the hot object into a lazily-created "callbacks" side object, so leaf nodes don't carry ~44 null refs.

• Optionally offer a lightweight leaf-variable type for the massive-tag-count case.

[marcschier]

We had a backlog item to page in nodestate when requested, i think this might be the fix. It was on the backlog as part of an improvement for redundent / shared state servers

[marcschier]

Note that the FolderState parent is only that large because it references all its members and its parent as well. So ultimately the entire memory allocation is the entire address space.

I am not sure if it makes sense to do anything urgently for 2.0, but we can look at a more lazy loading or "fault" based page in approach so that you get nodestate managed via GC and not retained in memory.