* perf(mito2): read compaction input footers concurrently
Compaction read the parquet metadata of its inputs one file at a time
before merging. It now reads up to 8 footers at once, keeping the input
order. Remote compaction also caches SST metadata now, so the merge
reuses those footers instead of reading each one a second time.
With 20ms object store reads, a remote compaction of 8 inputs issued 57
read requests instead of 65, and one output took 0.51s instead of 0.75s
with `experimental_compaction_output_parallelism` set to 8.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): stop accounting metadata bytes when caches drop
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* test(mito2): cover compaction with a missing input SST
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): skip output metadata warmup for remote compaction
Per-job compaction caches are not reused by readers of newly written SSTs. Keep input footer reuse, but preserve output prewarming only for long-lived engine caches. Cover both cache policies and the unchanged flush path with real SST reads.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
---------
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
When a datanode restarts with a new address, metasrv invalidated cached
table routes only for the tables in that heartbeat's region stats. The
first heartbeat after a restart often reports few or no regions (e.g.
with init_regions_in_background), so frontends kept dialing the old
address until metadata_cache_ttl expired.
Read the previous address via prev_kv when saving it, and when it
changed (or a failed save of the same epoch is retried) invalidate
every table registered for the datanode in __dn_table.
Signed-off-by: Floris Smit <florissmit10@gmail.com>
Co-authored-by: Weny Xu <wenymedia@gmail.com>
* feat(otlp): persist metric exemplars
OTLP metric conversion now emits exemplars through the shared storage contract. It validates timestamps, IDs, values and label budgets, and associates histogram exemplars only with emitted buckets that contain their values. Invalid observations are dropped and counted while valid points and siblings remain ingestible.
Frontend ingestion authorizes the emitted destinations and retains confirmed completions across metric and exemplar writes, including logical-table batching. Includes conversion and schema-failure isolation tests; tests that also query the new Prometheus endpoint land in the final layer.
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* feat(otlp): bind exemplars to source metric identities
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* fix(otlp): bound exemplar memory and preserve series
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* fix(otlp): skip exemplars without metric engine
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* fix(otlp): skip exemplars for ordinary tables
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* refactor(exemplar)!: prefix the shared table name
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
* fix(otlp): avoid waiting with body permits held
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
---------
Signed-off-by: shuiyisong <xixing.sys@gmail.com>
refactor(meta): separate migration construction and drain child tracking (#10098)
refactor(meta): separate migration construction and own drain child tracking
(cherry picked from commit 6a731a444f)
Signed-off-by: WenyXu <wenymedia@gmail.com>
Add a loader registration hook to TriggerDdlManager and invoke it
before procedure recovery. Test that registration reaches the
configured trigger DDL manager.
Signed-off-by: fys <fengys1996@gmail.com>
* feat: share streaming range results across scans
Share post-merge range batches across concurrent readers while preserving stream boundaries, cancellation and retry behavior. Encapsulate range ownership, compaction and publication in RangeResultCache, backed by the common SyncCache primitive.
Retain SST retirement fencing and memory accounting. Track live shared-buffer bytes and worker events alongside existing cache metrics.
Signed-off-by: WenyXu <wenymedia@gmail.com>
* fix(mito2): clean up shared range registrations on release
Signed-off-by: WenyXu <wenymedia@gmail.com>
---------
Signed-off-by: WenyXu <wenymedia@gmail.com>
DensePrimaryKeyCodec::decode_leftmost decoded the whole key and popped
the last value, so codecs with more than one field returned the
rightmost primary key column. Decode only the first field instead,
which also removes the full-key decoding noted in the TODO.
Current callers build single-field codecs and are unaffected.
Fixes#10089
Signed-off-by: Aumnertic <hznuzc@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
feat(postgres): OTLP binary payload ingest via COPY FROM STDIN
COPY FROM STDIN WITH (FORMAT otlp [, signal '...'] [, <options>]) ingests
one binary OTLP protobuf message over the postgres wire protocol — the
first binary COPY format (wire format code 1) — for all three signals:
- signal 'metrics' (default): ExportMetricsServiceRequest, completed
through OpenTelemetryProtocolHandler::metrics with the same context
preparation as the HTTP route (OtlpMetricCtx from the options); rows
are accepted data points.
- signal 'traces': ExportTraceServiceRequest through ::traces, with the
trace table name and pipeline name/version/params from the options
(an explicit pipeline name is required, exactly like the HTTP traces
endpoint); rows are accepted spans.
- signal 'logs': ExportLogsServiceRequest through ::logs, with the log
table name, pipeline options and extract keys from the options, and
the same OtlpLogDirect default as the HTTP route; rows are affected
rows.
The WITH options carry what the HTTP endpoints take as request headers,
reusing the HTTP path's own parsing helpers (truthy, SelectInfo,
GreptimePipelineParams, PipelineWay) so the two routes cannot drift;
values already provided by the postgres session (database, user,
permissions) are not repeated. Option semantics match the headers:
promote_all_resource_attrs / resource_attrs (';'-separated include or
exclude list) / promote_scope_attrs / translation_strategy for metrics;
table_name / pipeline_name / pipeline_version / pipeline_params /
extract_keys for traces and logs. Signal-inapplicable options are
rejected.
Like the influxdb format: table-less (a named table is rejected), the
whole payload is buffered until CopyDone exactly like one OTLP/HTTP
request body and bounded by the shared 64 MiB request limit, decoding
uses the same protobuf types, and the server must be wired —
PostgresServer::with_otlp_handler(handler, with_metric_engine) carries
the same metric-engine flag the HTTP wiring passes from its options.
The pipeline handler is derived from the protocol handler via trait
upcasting, as the HTTP routes do.
e2e: chunked binary metrics ingest; traces with table_name and
pipeline_name options; logs on the default table; metrics
promote_all_resource_attrs making a resource attribute queryable;
invalid-payload failure with connection reuse.
Signed-off-by: Ning Sun <sunning@greptime.com>
* fix(promql): collapse duplicate samples at one timestamp before range functions
A read path that does not deduplicate, such as an append-mode region, can
hand PromQL two rows for one series at one timestamp. Every range function
reads its samples from RangeManipulate, so the duplicate entries were
treated as independent samples: count_over_time counted both, idelta and
irate returned NULL because the last two entries shared a timestamp, and
delta/increase/rate deflated because the average sampling interval divides
the observed span by sample_count - 1.
Prometheus sample identity is one sample per series per timestamp, so its
range functions can never see this input. Restore that identity once, at
the operator every range function shares, instead of special-casing each
function's arithmetic.
The collapse compares native ticks, never the millisecond payload, so
distinct sub-millisecond samples that share a millisecond stay distinct as
calculate_range documents and tests. Batches without duplicate timestamps
take a single adjacency scan and are returned untouched; only when
duplicates exist is a take built, keeping tag and value columns aligned.
Keeping the last row of a group matches MergeMode::LastRow.
The normalize test that fed three rows at one timestamp now uses distinct
timestamps: it exercises histogram start-timestamp offsetting, and its
repeated timestamp was incidental, not asserted behavior.
Signed-off-by: discord9 <discord9@163.com>
* test(promql): cover duplicate samples end to end and keep the newer value
The operator now keeps the first row of a tied timestamp group, which is the row a
deduplicating read keeps: storage orders equal primary key and timestamp by sequence
descending, and FlatDedupReader takes the first row of each group. The end-to-end case
pins that choice by writing 100 and then 101 at the same timestamp: the window reports
increase 6, while keeping the stale 100 would report 8.
The sqlness case writes duplicate rows into an append_mode region, which is the read
path that does not deduplicate, and asserts every range function plus the unaffected
ordinary table in both the standalone and distributed environments.
Signed-off-by: discord9 <discord9@163.com>
* docs(promql): qualify duplicate-timestamp conflict selection
Signed-off-by: discord9 <discord9@163.com>
* perf(promql): avoid copying tags when collapsing duplicate samples
Signed-off-by: discord9 <discord9@163.com>
---------
Signed-off-by: discord9 <discord9@163.com>
* fix(datanode): pick a node epoch above the recorded one
The metasrv fences a datanode whose node epoch is lower than the
highest one it has recorded for the node id. A datanode used its start
time as its epoch, so after a restart on a host whose clock is behind
the previous process's, or after the clock stepped back, the new
process picked a lower epoch and was fenced for good: no region lease,
no keep-alive.
Before it creates the heartbeat task, the datanode now reads the
recorded epoch through its metadata backend and uses one more than it
if its start time is not higher. The node info still reports the local
start time. A failed read fails the startup, like the other metadata
reads.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* test(datanode): tighten the node epoch tests
Use a lower record that leaves room below the start time, so the test
fails if the start time is ignored, and a nonzero node id, so it fails
if the node id is ignored. Clarify the doc comment of the epoch choice.
This assumes the previous process of a node id has stopped before a new
one starts.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
---------
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* docs: link the observability database definition from README
Signed-off-by: Dennis Zhuang <killme2008@gmail.com>
* ci: add the new required checks to the Docs CI placeholders
The main ruleset now requires five more checks. Their workflows skip
docs-only PRs via paths-ignore, so without placeholders those PRs can
never satisfy the required checks.
Signed-off-by: Dennis Zhuang <killme2008@gmail.com>
---------
Signed-off-by: Dennis Zhuang <killme2008@gmail.com>
* feat: add relational region engine skeleton and registration
Add a read-only relational RegionEngine (create/open/drop lifecycle) and register it as the 4th engine in build_store_engines alongside mito/metric/file. Add a datanode-level test asserting all 4 engines are registered.
Signed-off-by: tian1220A <1573612521@qq.com>
* feat: add relational connection and idempotent region lifecycle
Move ConnectionInfo into relational-engine (DSN parsing + redacted rendering), place the RelationalOptions/RELATIONAL_META_KEY serialization contract in table::requests, make CREATE idempotent by comparing the persisted manifest definition against the incoming request (RegionExists on mismatch), and drop the empty EngineConfig placeholder.
Signed-off-by: tian1220A <1573612521@qq.com>
* refactor: address review comments on relational region engine
- remove connection.rs (only used by its own tests); it will be
reintroduced when connection handling is wired up
- drop RelationalOptions::version (not validated or used to select a
deserialization format)
- validate an idempotent create against the existing region definition
(region id / schema / primary key / options) and return Unexpected on
mismatch, with a reason field on RegionExists
- use serde.workspace and remove the empty [features] section
- cargo fmt
Signed-off-by: tian1220A <1573612521@qq.com>
* fix: adapt relational engine to latest main
- implement RegionEngine::set_region_role_with_lease (no-op lease, the
engine takes no writes so no lease is needed)
- pass node_id to build_object_store_manager in the registration test
- drop stale url dependency from the relational-engine lock entry
Signed-off-by: tian1220A <1573612521@qq.com>
* ci: register relational-engine in scope rules
Signed-off-by: tian1220A <1573612521@qq.com>
* fix: remove table self-dependency from Cargo.lock
Signed-off-by: tian1220A <1573612521@qq.com>
---------
Signed-off-by: tian1220A <1573612521@qq.com>
* perf(mito2): merge compaction inputs on a separate task
Compaction read, merged and encoded each output on a single task, so
parquet encoding waited for the merge and the merge waited for
encoding. The merge now runs on its own compaction runtime task and
hands merged batches to the SST writer through a bounded channel.
The channel also closes when that task panics, so the reading side
awaits the task before reporting the end of the input. A panic fails
the compaction instead of producing a truncated SST.
One output of 4.8M rows took 1.28s instead of 1.65s.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* perf(mito2): parallelize decoding and encoding of compaction outputs
Parquet encoding took about two thirds of a compaction output's time
and ran on a single thread. When `experimental_compaction_output_parallelism`
is above 1, the SST writer encodes the columns of each row group on up
to that many blocking threads of the compaction runtime while the next
row group is buffered, and the merge decodes up to that many input
files ahead of it on compaction runtime tasks.
The parallel writer feeds every column writer the same input as
`ArrowWriter`, so the SSTs are byte-identical. The bytes of a row group
reach the object store writer one row group later, so `max_file_size`
may split a file one row group later.
The option defaults to 1, which keeps today's CPU usage. One output of
4.8M rows took 0.44s with 4 and 0.38s with 8, against 1.28s with 1.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): cancel compaction producers before consumer polling
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* feat(mito2): support per-table compaction output parallelism
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): drain encoding tasks before returning failures
Abort remaining queued encoding tasks and await running tasks before propagating the first error, keeping row-group work inside the caller's compaction memory budget lifetime.
Add deterministic error and panic tests for queued cancellation, running-task cleanup, budget retention, and original error preservation.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): drain encoders on SST write errors
Add asynchronous abort cleanup to both SST write entry points so source errors cannot detach in-flight encoding work after its compaction memory budget is released. Share encoder task cleanup with the encoding-failure path and preserve the original write error.
Cover source-read failures in flat and primary-key writers with deterministic running and queued tasks.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(mito2): drain compaction outputs on cancellation
Stop scheduling new output batches on cancellation instead of aborting active writers. Await started outputs and retain successful SSTs, draining sibling tasks before propagating join failures.
Add deterministic cancellation regressions covering blocking encoder lifetime, budget retention, output ownership, join failures, and cancellation before work starts.
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
---------
Signed-off-by: Lei, HUANG <ratuthomm@gmail.com>
* fix(log-store): accept a retried first create of the conditional create probe
The startup probe of the object store WAL wrote its first create by hand
and failed on any error. When that create is stored but its response is
lost, the retry layer repeats it, the repeat meets the object, and the
probe reported a failed write, so the WAL refused to open.
Move the create of put_if_absent, which reads the object back and accepts
identical content, into a path-based core and use it for the probe's
first create too. The second create, which the store must refuse, is
unchanged.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* test(log-store): reuse LosingWriter for the retried probe create test
Move LosingWriter from the fault matrix into the test module and use it,
armed with the lose_next pattern, in place of RetriedWriter. Say why the
probe's second create bypasses put_path_if_absent: a read-back would find
the first create's object and let an overwriting store pass.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
---------
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* ci: gate the integration CI long tail by diff scope
`integration.yml` starts its long tail on every pull request: the three fuzz
families, the chaos legs, the export/import E2E job and the two sqlness
KvBackend legs. Most diffs cannot affect them, so a `scope` job now classifies
the pull request and tells those jobs whether to start.
The classifier is deterministic and fail-open. Forced modes (merge queue,
`schedule`, `workflow_dispatch`, `ci:full`, `CI_SCOPE_FORCE_ALL`) and three
floors -- always-run paths, an uncovered `src/` change, and a PromQL/flow
filter -- can only widen the result; everything else is a glob match against the
suite catalog. Every step of the job is `continue-on-error` and every gate reads
`needs.scope.outputs.<key> != 'false'`, so a classifier crash runs the jobs
instead of skipping them. The rules are checked out from the pull request base,
so a pull request cannot relax its own gate.
The policy is data: `.github/ci/scope-rules.toml` holds the path groups, the
suite catalog and the always-run list, with a `reason` on every exclusion.
`ci_scope_rules.py` loads it with the stdlib `tomllib` and validates the shape
(unknown or cyclic `@group` references, a `family` that disagrees with the suite
id, an empty reason) and that every glob's literal prefix still exists, so a
renamed crate cannot leave a rule that matches nothing. `make check-ci-scope`
runs that check locally; the `CI Scope Rules` job runs it on `ubuntu-slim`, the
same image the classifier uses. `.github/ci/scope-cases.toml` holds golden cases
that `test_ci_scope.py` replays through the real classifier.
The PostgreSQL and MySQL sqlness legs move into their own `sqlness-kvbackend`
job so they can be gated: a job-level `if:` cannot read the matrix, and gating
them inside `sqlness` would skip required checks and block the merge. They run
the same cases as the required Basic leg, so only the backend itself matters.
The verdict is visible where triage looks: one log line per suite, the exact
gate values, `::warning::` when a lookup fails and the gate widens, and a job
summary that lists each suite with the rule that decided it and the changed
files that matched.
Measured on 200 merged and open pull requests: 27.5% always run, 41% decided by
path rules, 31.5% skip every gated job; 995 of 2600 suite selections are
skipped, with no selection that a path rule makes. The classifier itself runs
in about ten seconds on a single-vCPU runner.
Signed-off-by: WenyXu <wenymedia@gmail.com>
* fix(ci): let the scope job read pull requests
The `scope` job declared `permissions: contents: read`, and an explicit block
sets every unlisted permission to `none`. `GET /pulls/{n}` therefore answered
403, `build_plan` widened to a full run as designed, and every gate read an
absent output as "run": the gate could never skip anything.
- `pull-requests: read` on the `scope` job.
- `tests/runner/**` joins `@kv-backend`: the runner implements
`--setup-pg`/`--setup-mysql`, which only the KV-backend legs pass, so a
change there must not skip the only job that exercises it. The
`UNMAPPED_CRATES` excuse is gone with it.
- The README no longer claims a queue failure blocks the merge: the gated jobs
are not required checks.
Signed-off-by: WenyXu <wenymedia@gmail.com>
* fix(ci): run every suite when the kind registry script changes
The distributed and chaos fuzz jobs start their cluster through
`.github/actions/setup-kind`, which runs `.github/scripts/kind-with-registry.sh`.
`.github/actions/**` is in `ALWAYS_RUN` but the script is not, so a change
confined to it selected no suite at all and every gate stayed `false`: the
jobs that exercise the changed cluster setup were skipped.
The script joins `ALWAYS_RUN`, consistent with the action that calls it, and a
golden case pins it.
Signed-off-by: WenyXu <wenymedia@gmail.com>
* fix(ci): widen the gate for the fuzz corpus seeding script
`fuzz-test/action.yaml` runs `.github/scripts/prepare-fuzz-corpus.sh` before
the cache restore in every fuzz job, but the script was not in `ALWAYS_RUN`.
A change confined to it therefore selected no suite: `any_fuzz` stayed
`false` and every fuzz job that runs it was skipped.
The script joins `ALWAYS_RUN` next to the other scripts the fuzz jobs execute,
with a golden case for it.
Both misses so far (this one and `kind-with-registry.sh`) came through a
composite action, which the workflow never names. `WorkflowContractTests`
now resolves `uses: ./.github/actions/<name>` transitively for the gated jobs
and fails when one of the scripts they run is not in `ALWAYS_RUN`.
Signed-off-by: WenyXu <wenymedia@gmail.com>
---------
Signed-off-by: WenyXu <wenymedia@gmail.com>
* fix: ignore stale leader grants after a region is downgraded
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix: compare stale leader grants with the longest lease seen
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix: treat a lease too long to store as unknown
A lease longer than a u64 of nanoseconds saturated to a shorter one, so the
derived send time could be later than the real one. Now such a lease makes
the send time unknown, which counts as earlier than any gate.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix: address review comments on stale leader grants
Scope the early send time claim to after a lease has been seen, drop the
unused role parameter of demote, clarify a test comment and check that a
fresh grant still makes a reopened region writable.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: compare stale leader grants with the heartbeat send time
Pass the send time of the heartbeat behind a lease to set_region_role_with_lease,
and return whether the role was applied. Mito compares the send time with the
grant gate directly, so it no longer derives it from the longest lease seen.
The datanode finalizes leadership only when the grant was applied.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* test: cover stale grants through the datanode and repeated downgrades
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: keep leader grant rules in one region method
Apply a role grant in MitoRegion::apply_role_grant, which renews the lease
and decides whether the role changes; the engine only looks up the region.
Rename the gate to downgrade_started_at and demote to begin_downgrade, keep
the field private, and copy it on reopen through
inherit_downgrade_state_from.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: drop a needless borrow
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: keep lease concepts in the engine and role ordering in the region
ManifestContext, which owns the role, records when the role was last
demoted and promotes a region only for a promotion issued after that. The
time starts when the context is created, so a reopened region ignores
grants for heartbeats sent before the reopen, without copying state. The
engine translates leases: it renews the deadline and promotes with the
heartbeat send time. MitoRegion keeps no grant state.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix: keep the data region of a metric region off a grant its metadata region ignores
The metric engine applies a grant to the metadata region first and stops
if it is ignored, so the data region, whose role the engine reports, never
becomes a leader that skips finalization. Also describe what promoting
returns.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* docs: describe when a lease grant is ignored and flatten the finalize check
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: move lease types out of the region and the Leader deadline into ManifestContext
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor: name the promotion fence after the downgrade it records
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
---------
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix(meta): register failure detectors for regions switched before a rollback
A region migration switches the routes of its tables one table at a
time. If it aborts after some tables were switched, those regions are
already served by the target, but the rollback deregistered the
target's failure detectors for every region and registered only the
source's. The target's next heartbeat registers them again, so if the
target stopped before sending one, nothing detected its failure and
the switched regions stayed unavailable until the metasrv leader
changed or the target came back.
The rollback now collects, from the routes it already reads, the
regions whose leader is the target and resets the target's failure
detectors for them at the end. It also skips a table whose route no
longer exists, so a table dropped meanwhile no longer fails the
rollback before the detectors are handled.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* test(meta): roll back a downgrading leader in the rollback detector test
The table that was not switched yet keeps a downgrading leader in a real
abort, so the test now builds it that way and checks that the rollback
clears it.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* fix(meta): monitor regions switched before a rollback only on the target
The rollback registered the source's failure detectors for every region
of the migration, including the regions already switched to the target.
Those regions are served by the target after the rollback, so their
source detectors watched the wrong peer: when one fired, the failover
found the leader had changed and only removed it. The rollback now
registers the source's detectors only for the regions that were not
switched, and keeps resetting the target's detectors for the others.
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
* refactor(meta): rename reset_failure_detectors_for_candidate_regions to reset_failure_detectors
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
---------
Signed-off-by: jeremyhi <fengjiachun@gmail.com>
Lei, HUANG
Floris SmitandWeny Xu
shuiyisong
LFC
Minh Vu
fys
liyang
AumandClaude Opus 5.5
Ning Sun
discord9
jeremyhi
dennis zhuang
zhaiAohan
Dragon Roar
Yvan Wang