Network presence push writes to one shared global table

[josephfusco]

#299 replaces the per-request UNION across each site's presence table with a push model: each site upserts its online-user row into one shared global table (wp_presence_network_summary), keyed by blog_id.

The write is coalesced, so a tick that finds the same people online is a zero-row no-op and the row is only rewritten when the user set changes or the row goes stale (TTL minus the idle heartbeat interval, 15s at the defaults). That bounds the rate but not the concentration.

Per-site tables can be split across database shards on a large network. A global table can't be split the same way, so every site's write still lands on one table instead of its own shard.

Does that write concentration hold up at large-network scale, or does the write path need to stay shardable?

PR: #299

[josephfusco]

No. The table registers with base_prefix into $wpdb->ms_global_tables, which is what a sharded network pins to the global cluster, while the per-site wp_presence tables carry the {base_prefix}{blog_id}_ shape a shard router places on the blog's own shard. The push moves presence writes off the shards onto the one cluster that cannot be split.

Rate is higher than I assumed when I filed this. wp_presence_enqueue_heartbeat_ping() writes the admin room on every admin pageview, not just the tick, and each write goes from 1 query to 4 with the push hooked in. Coalescing bounds row churn, not statement rate.

Fix is to gate the push on a change signal rather than run it on every write. wp_set_presence() is an upsert whose affected-row count already separates a new client row (1) from a refresh of one that was already there (2, or 0 when the row is byte-identical), and the delete paths know how many rows they removed, so an arrival or a departure is free to detect. Past that the only reason to push is to keep the row from aging out of the read cutoff at network-functions.php:345, which an autoloaded per-site option holding the last push time covers. Steady state then costs one push per site per wp_presence_network_summary_refresh_interval() instead of one per admin pageview, and needs no persistent object cache to get there.