Midhun Somasundaran
90d · built 2026-11-22
Performance
What Midhun Somasundaran shipped in the selected window, measured in ETV, and how it compares with the 90 days before it.
Effective capacity
+11.3engineers
delivers like 12.3 (12.3x pre-AI)
Output (ETV)
24.5ETV
+2396.9% vs 1.0 prior
Features share
26.3%
−17.6 pp vs prior window
Fixes share
34.1%
+16.8 pp vs prior window
Work mix
26.3% Features6.2% Maintenance32.8% Tests0.7% Docs34.1% Fixes
133 commits over 90 days, ending 2026-11-22.
Daily performance
Daily ETV, stacked by Features, Maintenance, Tests, Docs and Fixes.
Repository spread
Where this developer's commits land. Concentrated work (top1 > 80%) vs polymath spread (top1 < 30%).
Most impactful commits
Top 10 by ETV in the last 90 days.
- 1.4ETVApply VRF removals and interface VRF moves after route removal Summary: Config apply defers VRF removals and interface VRF moves until the routes using them are removed, so hardware never removes a VR or RIF that routes still use. Configs without such changes are applied as before. Config with VRF changes is applied in up to four steps: 1. Apply config with VRF removals and interface VRF moves deferred. Removed VRFs stay in state, and moving interfaces keep their old VRF, but their interface routes are left out of the RIB config. The final state is also validated here, before anything changes. 2. Program routes. The RIB removes all routes (from every client) and the FIB of removed VRFs, and the interface routes of moving interfaces from their old VRF. 3. Apply config normally: removed VRFs are deleted and interfaces move to their new VRF. 4. Program routes: interface routes of moved interfaces are added in their new VRF. Steps 3 and 4 run only if step 1 deferred something. Reviewed By: ravi861 Differential Revision: D127285001 fbshipit-source-id: ddafa74f4636ed4400f9211715d1433abda166dfgithub.com-facebook-fboss · be873e1d · 2026-11-03
- 0.9ETVProgram FIBs of several VRFs in one switch state update Summary: RibToSwitchStateFunction takes a list of VRF FIBs (RibVrfFib) instead of one VRF's tables, and RibToSwitchStateUpdater applies all of them in one state update. RibRouteTables::updateFib accepts several VRFs, each with its own deleted flag, and rolls back each of them on a HW update failure. Rollback rebuilds the global label FIB only into the default VRF; it used to be copied into whichever VRF was rolled back. Callers still pass one VRF. The linux agent callbacks handle every VRF they are given; a prefix is deleted from the kernel only when no VRF has it. The next diff uses this to program a VRF and the VRFs that resolve next hops in it together. Reviewed By: ravi861 Differential Revision: D127869157 fbshipit-source-id: d27f55f27f56f381ef779f607354cb94dbe22770github.com-facebook-fboss · ebfe5b1d · 2026-11-05
- 0.7ETVFix update thread deadlock in applyConfig ecmp override sync Summary: `applyConfigImpl` can deadlock the agent's update thread against concurrent route programming. After the config state update, applyConfig posted an ecmp override sync onto the update thread, and that task blocked on the rib thread. Meanwhile the rib thread, when it is programming a BGP route update, blocks on the update thread through `updateStateWithHwFailureProtection`. If a config apply lands while a route update is in its rib resolution phase, each thread waits on the other. This was confirmed from a core of a wedged agent: `fbossUpdateThread` was parked in `RoutingInformationBase::updateEcmpOverrides`, and `ribUpdateThread` was parked in `SwSwitch::updateStateBlockingImpl` from `RibRouteTables::updateFib`. Drains make it likely, because they restart bgpd right before reloading config. This is the inversion the comment above `routeUpdater.program()` already warns about. The override sync just never followed that rule. Fix: sync the overrides through the reconfigure workflow that `routeUpdater.program()` already uses. - `applyConfigImpl` no longer waits on the rib from the update thread. - `ThriftConfigApplier` records whether the backup switching mode changed (old state vs new config) on the `RouteUpdateWrapper` config routes. `program()` passes it to `RoutingInformationBase::reconfigure`, which runs on the rib thread, and `RibRouteTables::updateFib` then syncs overrides from all routes in the new state rather than only from the fib delta. The fib delta can't carry the change, because the config's own state update already re-moded the routes before this fib update. - That fib update compares the whole rib with the FIB, so it pushes the rib's backup-group routes still carrying the old mode. This relies on the parent diff, which makes EcmpResourceManager restamp those to the current mode; without it the old mode would revert the switch state before the sync. - `RoutingInformationBase::updateEcmpOverrides` had no other callers and is removed. Reviewed By: jasmeetbagga Differential Revision: D117291439 fbshipit-source-id: ea999efdaf15ad7e4aeb430e9568de146bb06f58github.com-facebook-fboss · a1563828 · 2026-10-12
- 0.7ETVAdd syncVrfs Thrift API Summary: Adds syncVrfs(list<VrfEntry>, map<string, i32> portNameToVrfId) to FbossCtrl and SwSwitch::syncVrfs(). In one switch state update it adds the VRFs not listed, moves the interfaces of the ports, renames VRFs, and deletes the VRFs not listed. Interfaces not listed go back to VRF 0. The VNI of an existing VRF cannot change. - The RIB computes the sync on its thread (RibRouteTables::syncVrfs), and the switch state change is computed from the latest state in the same HW-protected update, so a HW failure rolls back the RIB. A sync always runs the update. - updateInterfaceVrfs() and VRF adds and deletes share the same RIB and switch state code. A VRF deleted in the same update must have no interfaces after the moves, and only the routes of the interfaces that move out. - Marks restart event vrfs_synced on the first sync. Reviewed By: ravi861 Differential Revision: D132054386 fbshipit-source-id: 403a96c316724baf6089bd3f3f6551fe25edf9f2github.com-facebook-fboss · 4dc7150a · 2026-11-17
- 0.6ETVAdd resolvedInNextLevel to NextHop Summary: ResolvedNextHop may omit the interface for next hops resolved by a lookup in the next level. It stores resolvedInNextLevel, set when it has no interface, and a next hop is resolved when it has an interface or was resolved in the next level. Thrift encoding sets resolvedInNextLevel only for such next hops, so existing next hops encode as before, and old data where ifName means resolved still parses. Equality, ordering and hashes include it. No such next hop is created yet. Reviewed By: ravi861 Differential Revision: D128123908 fbshipit-source-id: 2bb64b691389503a788f20742c7f4c48b4605516github.com-facebook-fboss · 99faaba1 · 2026-11-04
- 0.6ETVReject resolvedInNextLevel in client next hops Summary: Only the agent sets resolvedInNextLevel on resolved next hops. Reject it in client next hops: unicast, backup, MPLS, MySid and named next hop group APIs, and static MPLS and IP to MPLS routes in config. Reviewed By: ravi861 Differential Revision: D128190268 fbshipit-source-id: 35170cef8d1b7632f63862bc29d0046360053717github.com-facebook-fboss · b1ccbc70 · 2026-11-04
- 0.6ETVSaiSwitch programs local router interface VRFs from the interface VRF map Summary: SaiSwitch passes each local router interface's VRF from the interface VRF map to SaiRouterInterfaceManager and its to-me routes, and moves a router interface when only its map entry changes. Remote router interfaces keep the VRF from their peer. The delta order is unchanged. Reviewed By: jasmeetbagga Differential Revision: D130468412 fbshipit-source-id: 462fb8209afefae40afc8821fa83c64f7ad34fa7github.com-facebook-fboss · cb0a016d · 2026-11-16
- 0.5ETVConfig apply takes interface VRFs from the interface VRF map Summary: Config apply takes an interface's VRF from its interface VRF map entry, else from config routerID. It rejects removing an interface or VRF that a map entry still uses. addVrfForPorts() no longer sets routerID. AgentVrfTest and AgentVxlanTests bind their VRF ports with updateInterfaceVrfs() in cold boot setup; the config move and delete tests still bind through config routerID. Reviewed By: tanmayaudupa Differential Revision: D130438838 fbshipit-source-id: 2c3dccc4216e956413d313777e877aaeef8aa0e5github.com-facebook-fboss · e76355c4 · 2026-11-13
- 0.5ETVResolve interface-scoped next hops only in their interface's VRF Summary: On switches with VRFs, a client route's interface-scoped (link-local) next hop resolves only if its interface belongs to the route's VRF. Interfaces moving between VRFs belong to no VRF until the move, so routes through them are removed before the router interface moves. Switches without VRFs are unchanged. Reviewed By: ravi861 Differential Revision: D127303684 fbshipit-source-id: 3f5d3239f7bb537994731e7e926ecd9c921104e5github.com-facebook-fboss · a72b9db2 · 2026-11-03
- 0.5ETVFlush dynamic FDB entries when a port is MAC-limit error-disabled Summary: When a port newly carries `ERROR_DISABLE_MAC_LIMIT_EXCEEDED`, SaiSwitch flushes the port's dynamic FDB entries after the port admin disable, in the same state delta. This frees L2 table space held by spurious MACs, including entries learned in HW that never reached FBOSS. - The MACs the delta removes from the port are deleted from HW as usual, before the port is disabled. On TH3 and newer, FBOSS never deletes dynamic entries one by one (they're created skip-remove, and the SDK owns them), so the flush is what removes them there. Anything the port learns between those removals and the disable is also cleared by the flush. - Nothing extra is marked skip-remove for the flush. If the flush can't run (for example the port has no bridge port), HW isn't left holding entries FBOSS stopped tracking because it expected the flush to remove them. - The flush runs after the bridge ports are added. Treating added ports that carry the error as newly disabled re-flushes them on warm boot, covering a restart between the state commit and the flush. - The flush is not undone on rollback; the entries are dynamic. A failed flush is logged and doesn't fail the update, so the port still goes down. - Only ASICs with the new `HwAsic::Feature::PORT_MAC_FLUSH` (TH, TH3-TH6) flush; elsewhere, removing the MACs from the switch state removes them from the SDK. Reviewed By: tanmayaudupa Differential Revision: D125375536 fbshipit-source-id: 893856e2e32f07d80f4fe9f82e2dfc674bd4f3cagithub.com-facebook-fboss · 0b494d75 · 2026-11-18