| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
fix[faasscheduler]: add session-based lightweight scheduling branch (LiteScheduler) with noopSender Add a session-sticky lightweight scheduling bypass (LiteScheduler) on top of the existing FaaSScheduler. When a request carries a sessionID, it enters the lite branch instead of the legacy PoolManager/GenericInstancePool path. The lite branch provides acquire/release/retain/batchRetain with bounded cold-start polling and per-request Prometheus metrics. Cold-start design: LiteScheduler constructs a ScaleHint (capacity demand hint) and sends it via the ScaleHintSender interface. The current default implementation is noopSender (logs only, no scaler dispatch), which serves as a placeholder for future HTTP/gRPC sender implementations. Instance creation is handled by ReplicaScaler's minInstance mechanism. The ScaleHintSender interface is the single seam: swapping noopSender for an httpSender/grpcSender requires no changes to operation.go or litescheduler.go. Architecture: Core types (go/pkg/functionscaler/): - types.go: LiteSchedulerConfig (Enable/EnableAllTenants/EnabledTenants/ EnabledFunctions/AcquireWaitTimeoutMs/UseExistingLease/SessionIdleTimeoutS) - config.go: normalizeLiteScheduler fills defaults (timeout<=0 -> 3000, nil slices -> empty) - selfregister/proxy.go: CheckHashOwner generalizes CheckFuncOwner to an arbitrary hash key (funcKey for legacy, tenantID/sessionID for lite). - scaler/instance_scaler.go: ScaleHint struct (no ReceiveScaleHint on the InstanceScaler interface; noopSender replaces inProcessSender). - faasscheduler.go: mount LiteScheduler (enable-guarded), bypass entry before the legacy switch; inject NewNoopSender(); register LiteCollector with the default prometheus registry. - metrics/prometheus_metric.go: relax InitServerMetric guard from ScenarioWiseCloud-only to MetricsAddr-non-empty. New package go/pkg/functionscaler/litescheduler/: - allocation.go: Allocation + allocationID gen/parse (lite:{sha256-16hex}:{instanceID}:thread:{seq}; full sessionID not embedded to avoid leakage). - pool.go: LiteInstance/LiteFunctionPool/PoolStats/Dispatcher interface + status mapping (Running/SubHealth/Unavailable) + Stats()/candidateSlotsLocked. - dispatcher.go: concurrency (lowest InUse/Capacity + subHealth penalty) and round-robin (cursor, healthy-then-subHealth fallback) dispatchers; newDispatcher picks by SchedulePolicy, unknown degrades to concurrency. - litescheduler.go: LiteScheduler (pools/allocations + RWMutex, ownerProxy, funcSpecGetter, scaleHintSender, stopCh) + New (live-read config, no snapshot) + isFuncEnabled three-tier whitelist + Pools()/Metrics() accessors. - request.go: LiteRequest + ParseRequest trigger matrix (6 paths, ok=false falls back to legacy) + extractSessionID + panic recovery. - operation.go: handleAcquire (session-sticky + dispatch + cold-start) / assignInstance / handleRelease / handleRetain / handleBatchRetain / handleColdStart (ScaleHint + bounded waitForInstance polling) with pool.Lock->allocMu.Lock single-direction lock ordering (no AB-BA). - event.go: three independent registry subscriptions (FuncSpec/InsSpec/ SchedulerProxy) + goroutine loops with select stopCh for graceful exit; upsertPool/deletePool/handleInstanceUpdate handleInstanceDelete keep instances/sessions/allocations in sync. - scalehint.go: noopSender logs hints but does not dispatch them; future implementations (httpSender/grpcSender) can be injected via New(). - metrics.go: LiteCollector (8 faas_lite_* metrics, read-at-scrape Gauges + request-path Counters); incAcquire/incRelease/incRetain/incScaleHint wired into operation.go handlers. - Process: reverse lookup (alloc -> sessionID/tenantID/funcKey), isFuncEnabled recheck, owner check (acquire-only, traffic owner = tenantID/sessionID), dispatch with panic recovery. Deploy rendering: - go/build/faas/init_scheduler_args.json (liteScheduler block with {liteXxx} placeholders) + api/python/yr/cli/values.toml ([values.lite_scheduler]) + faas_scheduler.py/function_scheduler.py. - K8s: deploy/k8s/charts/openyuanrong values.yaml (global.functionScheduler. liteScheduler) + scheduler-configmap.yaml ({{ ... | toJson }}). Logging: - Structured logs (Info/Warn/Debug/Error with traceID/funcKey/sessionID fields) across all litescheduler paths. - Log-side snapshots fixed three latent data races exposed by reading pool/alloc fields after Unlock: handleAcquire cold-start instanceCount, handleRelease inUse/capacity, handleRetain newExpire. -race clean. Testing: - TDD across all tasks: failing test -> implementation -> green. - litescheduler package: -race clean (no data race, no deadlock). - Coverage: allocationID format/parse edge cases, status mapping, dispatcher concurrency/round-robin/unavailable exclusion, isFuncEnabled three-tier, ParseRequest trigger matrix, acquire/release/retain/batchRetain semantics, cold-start bounded wait + concurrent map-race guard, Process reverse lookup/owner check, registry event pool maintenance, metrics inc-wired, noopSender nil-hint no-panic. Verified with request da2190af-dace-489e-8fe3-b2e1cae8e365: - LiteScheduler correctly enters cold-start path, constructs ScaleHint, sends via noopSender (logged), and waits for instance via polling. - Instance creation handled by ReplicaScaler minInstance mechanism. - batchRetain chain works correctly with TTL refresh. - Known issue: release does not reach lite branch (frontend handles locally and discards lite-prefixed leaseId), causing InUse count leak. Fix requires frontend lease_manager changes (out of scope). Signed-off-by: For_YL <zhangtangwei@huawei.com> | 1 天前 | |
chore: consolidate sandbox post-merge fixes Preserve the original sandbox merge history while folding the follow-up CI, codecheck, packaging, deploy, and conflict-resolution commits into one merge commit. The tree is identical to the previous MR head 25a7b4e1fd7ea62d90e2e310a9b7414e4fb2d8a4. Signed-off-by: mhsong2 <songminhui2@huawei.com> | 1 个月前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
feat(trace): optimize faas trace propagation Signed-off-by: Lwy_Robb <luowenyu4@huawei.com> | 3 个月前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
chore: consolidate sandbox post-merge fixes Preserve the original sandbox merge history while folding the follow-up CI, codecheck, packaging, deploy, and conflict-resolution commits into one merge commit. The tree is identical to the previous MR head 25a7b4e1fd7ea62d90e2e310a9b7414e4fb2d8a4. Signed-off-by: mhsong2 <songminhui2@huawei.com> | 1 个月前 | |
feat[scheduler]: move ScaleHint struct from scaler to litescheduler package ScaleHint is a LiteScheduler concept (capacity demand hint) that was defined in the scaler package but never consumed by any scaler code after ReceiveScaleHint was removed. Moving it to litescheduler/scalehint.go places it with its sole consumer and eliminates the litescheduler → scaler import dependency in 4 files (scalehint.go, litescheduler.go, operation.go, scalehint_test.go). add session ttl unbind operation Signed-off-by: For_YL <zhangtangwei@huawei.com> | 1 天前 | |
add license | 7 个月前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
fix(ci): support runtime manager packaging Signed-off-by: mhsong2 <songminhui2@huawei.com> | 11 天前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 | |
fix[scheduler]: 修复 SchedulerProxy.Add 签名变更后 proxy_test 编译失败 - 引入 LiteScheduler 后 SchedulerProxy.Add 新增 tokenType、currentVersionFlag 参数 - proxy_test.go 中 sp.Add 调用仍为旧 2 参数，导致测试编译失败 - 同步更新调用为 Add(info, "", "", true)，恢复编译 Signed-off-by: For_YL <zhangtangwei@huawei.com> | 18 小时前 | |
feat[all]: merge ant branch with sandbox, snapshot, and aio features Major changes: - Add sandbox functionality for function execution isolation - Add snapshot/snapstart support for checkpoint management - Add AIO (all-in-one) example with Docker deployment - Support loading environment variables from .env files - Add port forwarding functionality - Improve multi-language APIs (Python/Java/Go/C++) - Optimize build and deployment scripts - Update documentation and examples - Add submodules: datasystem, frontend, functionsystem Signed-off-by: songminhui <songminhui2@huawei.com> | 3 个月前 | |
modify go moudle name | 7 个月前 | |
modify go moudle name | 7 个月前 | |
build: 增加faascheduler和faasmanager的llt脚本 Signed-off-by: akabanekun <taokun3@h-partners.com> | 3 个月前 | |
chore: consolidate sandbox post-merge fixes Preserve the original sandbox merge history while folding the follow-up CI, codecheck, packaging, deploy, and conflict-resolution commits into one merge commit. The tree is identical to the previous MR head 25a7b4e1fd7ea62d90e2e310a9b7414e4fb2d8a4. Signed-off-by: mhsong2 <songminhui2@huawei.com> | 1 个月前 | |
fix[faasscheduler]: add session-based lightweight scheduling branch (LiteScheduler) with noopSender Add a session-sticky lightweight scheduling bypass (LiteScheduler) on top of the existing FaaSScheduler. When a request carries a sessionID, it enters the lite branch instead of the legacy PoolManager/GenericInstancePool path. The lite branch provides acquire/release/retain/batchRetain with bounded cold-start polling and per-request Prometheus metrics. Cold-start design: LiteScheduler constructs a ScaleHint (capacity demand hint) and sends it via the ScaleHintSender interface. The current default implementation is noopSender (logs only, no scaler dispatch), which serves as a placeholder for future HTTP/gRPC sender implementations. Instance creation is handled by ReplicaScaler's minInstance mechanism. The ScaleHintSender interface is the single seam: swapping noopSender for an httpSender/grpcSender requires no changes to operation.go or litescheduler.go. Architecture: Core types (go/pkg/functionscaler/): - types.go: LiteSchedulerConfig (Enable/EnableAllTenants/EnabledTenants/ EnabledFunctions/AcquireWaitTimeoutMs/UseExistingLease/SessionIdleTimeoutS) - config.go: normalizeLiteScheduler fills defaults (timeout<=0 -> 3000, nil slices -> empty) - selfregister/proxy.go: CheckHashOwner generalizes CheckFuncOwner to an arbitrary hash key (funcKey for legacy, tenantID/sessionID for lite). - scaler/instance_scaler.go: ScaleHint struct (no ReceiveScaleHint on the InstanceScaler interface; noopSender replaces inProcessSender). - faasscheduler.go: mount LiteScheduler (enable-guarded), bypass entry before the legacy switch; inject NewNoopSender(); register LiteCollector with the default prometheus registry. - metrics/prometheus_metric.go: relax InitServerMetric guard from ScenarioWiseCloud-only to MetricsAddr-non-empty. New package go/pkg/functionscaler/litescheduler/: - allocation.go: Allocation + allocationID gen/parse (lite:{sha256-16hex}:{instanceID}:thread:{seq}; full sessionID not embedded to avoid leakage). - pool.go: LiteInstance/LiteFunctionPool/PoolStats/Dispatcher interface + status mapping (Running/SubHealth/Unavailable) + Stats()/candidateSlotsLocked. - dispatcher.go: concurrency (lowest InUse/Capacity + subHealth penalty) and round-robin (cursor, healthy-then-subHealth fallback) dispatchers; newDispatcher picks by SchedulePolicy, unknown degrades to concurrency. - litescheduler.go: LiteScheduler (pools/allocations + RWMutex, ownerProxy, funcSpecGetter, scaleHintSender, stopCh) + New (live-read config, no snapshot) + isFuncEnabled three-tier whitelist + Pools()/Metrics() accessors. - request.go: LiteRequest + ParseRequest trigger matrix (6 paths, ok=false falls back to legacy) + extractSessionID + panic recovery. - operation.go: handleAcquire (session-sticky + dispatch + cold-start) / assignInstance / handleRelease / handleRetain / handleBatchRetain / handleColdStart (ScaleHint + bounded waitForInstance polling) with pool.Lock->allocMu.Lock single-direction lock ordering (no AB-BA). - event.go: three independent registry subscriptions (FuncSpec/InsSpec/ SchedulerProxy) + goroutine loops with select stopCh for graceful exit; upsertPool/deletePool/handleInstanceUpdate handleInstanceDelete keep instances/sessions/allocations in sync. - scalehint.go: noopSender logs hints but does not dispatch them; future implementations (httpSender/grpcSender) can be injected via New(). - metrics.go: LiteCollector (8 faas_lite_* metrics, read-at-scrape Gauges + request-path Counters); incAcquire/incRelease/incRetain/incScaleHint wired into operation.go handlers. - Process: reverse lookup (alloc -> sessionID/tenantID/funcKey), isFuncEnabled recheck, owner check (acquire-only, traffic owner = tenantID/sessionID), dispatch with panic recovery. Deploy rendering: - go/build/faas/init_scheduler_args.json (liteScheduler block with {liteXxx} placeholders) + api/python/yr/cli/values.toml ([values.lite_scheduler]) + faas_scheduler.py/function_scheduler.py. - K8s: deploy/k8s/charts/openyuanrong values.yaml (global.functionScheduler. liteScheduler) + scheduler-configmap.yaml ({{ ... | toJson }}). Logging: - Structured logs (Info/Warn/Debug/Error with traceID/funcKey/sessionID fields) across all litescheduler paths. - Log-side snapshots fixed three latent data races exposed by reading pool/alloc fields after Unlock: handleAcquire cold-start instanceCount, handleRelease inUse/capacity, handleRetain newExpire. -race clean. Testing: - TDD across all tasks: failing test -> implementation -> green. - litescheduler package: -race clean (no data race, no deadlock). - Coverage: allocationID format/parse edge cases, status mapping, dispatcher concurrency/round-robin/unavailable exclusion, isFuncEnabled three-tier, ParseRequest trigger matrix, acquire/release/retain/batchRetain semantics, cold-start bounded wait + concurrent map-race guard, Process reverse lookup/owner check, registry event pool maintenance, metrics inc-wired, noopSender nil-hint no-panic. Verified with request da2190af-dace-489e-8fe3-b2e1cae8e365: - LiteScheduler correctly enters cold-start path, constructs ScaleHint, sends via noopSender (logged), and waits for instance via polling. - Instance creation handled by ReplicaScaler minInstance mechanism. - batchRetain chain works correctly with TTL refresh. - Known issue: release does not reach lite branch (frontend handles locally and discards lite-prefixed leaseId), causing InUse count leak. Fix requires frontend lease_manager changes (out of scope). Signed-off-by: For_YL <zhangtangwei@huawei.com> | 1 天前 | |
chore: Merge upstream master into sandbox merge branch Signed-off-by: mhsong2 <songminhui2@huawei.com> | 19 天前 | |
chore: consolidate sandbox post-merge fixes Preserve the original sandbox merge history while folding the follow-up CI, codecheck, packaging, deploy, and conflict-resolution commits into one merge commit. The tree is identical to the previous MR head 25a7b4e1fd7ea62d90e2e310a9b7414e4fb2d8a4. Signed-off-by: mhsong2 <songminhui2@huawei.com> | 1 个月前 | |
fix[faasscheduler]: add session-based lightweight scheduling branch (LiteScheduler) with noopSender Add a session-sticky lightweight scheduling bypass (LiteScheduler) on top of the existing FaaSScheduler. When a request carries a sessionID, it enters the lite branch instead of the legacy PoolManager/GenericInstancePool path. The lite branch provides acquire/release/retain/batchRetain with bounded cold-start polling and per-request Prometheus metrics. Cold-start design: LiteScheduler constructs a ScaleHint (capacity demand hint) and sends it via the ScaleHintSender interface. The current default implementation is noopSender (logs only, no scaler dispatch), which serves as a placeholder for future HTTP/gRPC sender implementations. Instance creation is handled by ReplicaScaler's minInstance mechanism. The ScaleHintSender interface is the single seam: swapping noopSender for an httpSender/grpcSender requires no changes to operation.go or litescheduler.go. Architecture: Core types (go/pkg/functionscaler/): - types.go: LiteSchedulerConfig (Enable/EnableAllTenants/EnabledTenants/ EnabledFunctions/AcquireWaitTimeoutMs/UseExistingLease/SessionIdleTimeoutS) - config.go: normalizeLiteScheduler fills defaults (timeout<=0 -> 3000, nil slices -> empty) - selfregister/proxy.go: CheckHashOwner generalizes CheckFuncOwner to an arbitrary hash key (funcKey for legacy, tenantID/sessionID for lite). - scaler/instance_scaler.go: ScaleHint struct (no ReceiveScaleHint on the InstanceScaler interface; noopSender replaces inProcessSender). - faasscheduler.go: mount LiteScheduler (enable-guarded), bypass entry before the legacy switch; inject NewNoopSender(); register LiteCollector with the default prometheus registry. - metrics/prometheus_metric.go: relax InitServerMetric guard from ScenarioWiseCloud-only to MetricsAddr-non-empty. New package go/pkg/functionscaler/litescheduler/: - allocation.go: Allocation + allocationID gen/parse (lite:{sha256-16hex}:{instanceID}:thread:{seq}; full sessionID not embedded to avoid leakage). - pool.go: LiteInstance/LiteFunctionPool/PoolStats/Dispatcher interface + status mapping (Running/SubHealth/Unavailable) + Stats()/candidateSlotsLocked. - dispatcher.go: concurrency (lowest InUse/Capacity + subHealth penalty) and round-robin (cursor, healthy-then-subHealth fallback) dispatchers; newDispatcher picks by SchedulePolicy, unknown degrades to concurrency. - litescheduler.go: LiteScheduler (pools/allocations + RWMutex, ownerProxy, funcSpecGetter, scaleHintSender, stopCh) + New (live-read config, no snapshot) + isFuncEnabled three-tier whitelist + Pools()/Metrics() accessors. - request.go: LiteRequest + ParseRequest trigger matrix (6 paths, ok=false falls back to legacy) + extractSessionID + panic recovery. - operation.go: handleAcquire (session-sticky + dispatch + cold-start) / assignInstance / handleRelease / handleRetain / handleBatchRetain / handleColdStart (ScaleHint + bounded waitForInstance polling) with pool.Lock->allocMu.Lock single-direction lock ordering (no AB-BA). - event.go: three independent registry subscriptions (FuncSpec/InsSpec/ SchedulerProxy) + goroutine loops with select stopCh for graceful exit; upsertPool/deletePool/handleInstanceUpdate handleInstanceDelete keep instances/sessions/allocations in sync. - scalehint.go: noopSender logs hints but does not dispatch them; future implementations (httpSender/grpcSender) can be injected via New(). - metrics.go: LiteCollector (8 faas_lite_* metrics, read-at-scrape Gauges + request-path Counters); incAcquire/incRelease/incRetain/incScaleHint wired into operation.go handlers. - Process: reverse lookup (alloc -> sessionID/tenantID/funcKey), isFuncEnabled recheck, owner check (acquire-only, traffic owner = tenantID/sessionID), dispatch with panic recovery. Deploy rendering: - go/build/faas/init_scheduler_args.json (liteScheduler block with {liteXxx} placeholders) + api/python/yr/cli/values.toml ([values.lite_scheduler]) + faas_scheduler.py/function_scheduler.py. - K8s: deploy/k8s/charts/openyuanrong values.yaml (global.functionScheduler. liteScheduler) + scheduler-configmap.yaml ({{ ... | toJson }}). Logging: - Structured logs (Info/Warn/Debug/Error with traceID/funcKey/sessionID fields) across all litescheduler paths. - Log-side snapshots fixed three latent data races exposed by reading pool/alloc fields after Unlock: handleAcquire cold-start instanceCount, handleRelease inUse/capacity, handleRetain newExpire. -race clean. Testing: - TDD across all tasks: failing test -> implementation -> green. - litescheduler package: -race clean (no data race, no deadlock). - Coverage: allocationID format/parse edge cases, status mapping, dispatcher concurrency/round-robin/unavailable exclusion, isFuncEnabled three-tier, ParseRequest trigger matrix, acquire/release/retain/batchRetain semantics, cold-start bounded wait + concurrent map-race guard, Process reverse lookup/owner check, registry event pool maintenance, metrics inc-wired, noopSender nil-hint no-panic. Verified with request da2190af-dace-489e-8fe3-b2e1cae8e365: - LiteScheduler correctly enters cold-start path, constructs ScaleHint, sends via noopSender (logged), and waits for instance via polling. - Instance creation handled by ReplicaScaler minInstance mechanism. - batchRetain chain works correctly with TTL refresh. - Known issue: release does not reach lite branch (frontend handles locally and discards lite-prefixed leaseId), causing InUse count leak. Fix requires frontend lease_manager changes (out of scope). Signed-off-by: For_YL <zhangtangwei@huawei.com> | 1 天前 | |
feat: align scheduler routing changes Sync the January fork scheduler routing changes into the current repository so function invocation can use the blue-green ring ownership model instead of the old gray rollout path. This removes the obsolete rollout self-register flow, updates scheduler registry and routing logic, and makes non-owner schedulers return owner information for retry and switch handling. The runtime side adds scheduler address management and cache support so invoke requests can locate the correct scheduler during direct invocation and function intercall scenarios. Review fixes are included for rollout ratio locking and reset, consistent ring boundary checks, scheduler ownership state, non-owner retry keys, batch renew routing, and related codecheck findings. Tests are updated for scheduler registry, concurrency/session scheduling, runtime scheduler manager, lease pending request handling, rollout ratio deletion, and related invocation paths to cover the synchronized behavior. Signed-off-by: neatline <fuyipeng3@huawei.com> | 3 天前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 1 天前 | ||
| 1 个月前 | ||
| 3 天前 | ||
| 3 个月前 | ||
| 3 天前 | ||
| 3 天前 | ||
| 1 个月前 | ||
| 1 天前 | ||
| 7 个月前 | ||
| 3 天前 | ||
| 11 天前 | ||
| 3 天前 | ||
| 3 天前 | ||
| 3 天前 | ||
| 18 小时前 | ||
| 3 个月前 | ||
| 7 个月前 | ||
| 7 个月前 | ||
| 3 个月前 | ||
| 1 个月前 | ||
| 1 天前 | ||
| 19 天前 | ||
| 1 个月前 | ||
| 1 天前 | ||
| 3 天前 |