[em3s]

Summary

JdbcHashLabel fetches graph.metadata-fetch-limit rows per scan key and filters inactive ones out afterwards. A metadata DELETE only tombstones, so tombstones spend that budget, and past the limit the page is truncated silently — AbstractLabel.scan computes hasNext and DdlService.getAll drops it.

EdgeIndex on v3 is the fix. Until then I needed to see how close deployed servers are, so this adds a command that measures it.

jdbc-metastore remaining walks /graph/v2/metastore/global — the only endpoint that shows the rows a scan reads, tombstones included — and counts per (src, labelId), the scan key getAll builds. Both values come off the server's decode, so nothing re-implements the codec. The endpoint is 0.4.x only, removed by #418, so the command probes first and says so.

No server change. Same tombstone problem as #485.

Notes

• --capacity is required: it is the server's graph.metadata-fetch-limit, deployments override the OSS default of 1000, and actuator masks the value.
• Every window prints, grouped by namespace, since a metastore table can be shared between deployments.
• Bounded: counters per page, 500-row pages, five stop conditions (last, empty page, --max-rows, --max-pages, total reached). Whichever cap fires is named and exits non-zero. Ctrl-C reports what it counted.
• Not using the server's prefix= count: buildQuery builds k LIKE '<prefix>%' unescaped, and base64url keys contain _.

Changes

• Add jdbc-metastore remaining, with --capacity, --page-size, --max-pages, --max-rows
• Add GetStream to the HTTP client, which keeps the real status code and decodes off the connection
• Add DumpMetastore(page, size) and the dump page model
• Register the command

How to Test

cd cli && go test ./internal/command/ && go build ./... && go vet ./...

Against a 0.4.x server. Lower both to see the classification without a large table:

actionbase --host http://<0.4.x-host>:8080 -e "jdbc-metastore remaining --capacity 3 --page-size 2"

20 rows, capacity 3/window, pages of 2 (max 2000). Ctrl-C reports partial.
alpha         14 rows  4 windows
  service      OVER by 2   alpha:origin               5/3     live 5      tomb 0
  storage      OVER by 2   alpha:origin               5/3     live 1      tomb 4
  label        1 left      alpha:talk                 2/3     live 0      tomb 2
               1 left      alpha:wish                 2/3     live 0      tomb 2
beta           5 rows  2 windows
  service      2 left      beta:origin                1/3     live 1      tomb 0
  label        OVER by 1   beta:gift                  4/3     live 2      tomb 2
1 undecodable (window unknown, e.g. k="not-an-encoded-key")
20/20 rows, 6 windows, 2 namespaces. 3 over, 0 near. An over-capacity window is already truncated and does not say so.

All of these exit non-zero, as does a window over capacity:

4/20 rows, 1 window, 1 namespace. Stopped at --max-pages 2 - incomplete, 1 over is a lower bound.
--capacity is required: the server's graph.metadata-fetch-limit, which it does not expose (actuator masks it). OSS default 1000.
No /graph/v2/metastore/global on this server (0.4.x only), so this cannot be measured here.
Unauthorized. Set --authKey or ACT_API_KEY.

Sample output and the last two paths came from a stub server; the rest ran against real 0.4.x servers.

Not touched

The root cause, which EdgeIndex addresses. Also MetastoreInspector.buildQuery's _ escaping, which is on 0.4.x and below rather than main, and CheckConnection exiting 0 on an unreachable host, which predates this and affects every command.

AI Assistance

• This PR was written largely with AI assistance.
  • Tool / model: Claude Code (Opus 5)

[em3s]

Optimistic