store-gateways crash on startup if they encounter intermittent error

[gouthamve]

Describe the bug
store-gateways do a LIST operation to get all the users and then load the bucket index for that user. These operations might hit a temporary timeout and this causes the store-gateway to crash and restart again. We should be resilient against minor issues like this.

level=warn ts=2021-02-09T10:53:50.301491118Z caller=bucket_stores.go:128 msg="failed to synchronize TSDB blocks" err="failed to synchronize TSDB blocks for user 198923: sync block: read bucket index: read bucket index: Get \"https://storage.googleapis.com/<bucket>/198923/bucket-index.json.gz\": net/http: TLS handshake timeout"
level=info ts=2021-02-09T10:53:50.301685441Z caller=basic_lifecycler.go:197 msg="ring lifecycler is shutting down" ring=store-gateway
level=error ts=2021-02-09T10:53:50.301711626Z caller=client.go:233 msg="error getting path" key=store-gateway err="Get \"http://consul.cortex-blocks-prod.svc.cluster.local:8500/v1/kv/store-gateway?index=73476708&stale=&wait=10000ms\": context canceled"
level=info ts=2021-02-09T10:53:50.310302637Z caller=basic_lifecycler.go:363 msg="unregistering instance from ring" ring=store-gateway
level=info ts=2021-02-09T10:53:50.320575066Z caller=basic_lifecycler.go:234 msg="instance removed from the ring" ring=store-gateway
level=error ts=2021-02-09T10:53:50.320726791Z caller=cortex.go:402 msg="module failed" module=store-gateway err="invalid service state: Failed, expected: Running, failure: initial blocks synchronization: failed to synchronize TSDB blocks for user 198923: sync block: read bucket index: read bucket index: Get \"https://storage.googleapis.com/<bucket>/198923/bucket-index.json.gz\": net/http: TLS handshake timeout"
level=info ts=2021-02-09T10:53:50.320787529Z caller=module_service.go:90 msg="module stopped" module=runtime-config
level=info ts=2021-02-09T10:53:50.320854193Z caller=module_service.go:90 msg="module stopped" module=memberlist-kv
level=info ts=2021-02-09T10:53:50.321051245Z caller=server_service.go:50 msg="server stopped"
level=info ts=2021-02-09T10:53:50.321073044Z caller=module_service.go:90 msg="module stopped" module=server
level=info ts=2021-02-09T10:53:50.321088676Z caller=cortex.go:391 msg="Cortex stopped"
level=error ts=2021-02-09T10:53:50.321159052Z caller=log.go:27 msg="error running cortex" err="failed services\ngithub.com/cortexproject/cortex/pkg/cortex.(*Cortex).Run...."

Storage Engine

• Blocks
• Chunks

[pracucci]

As easy improvement would be adding retries to the per-user blocks synchronization. We did something similar in the compactor:

cortex/pkg/compactor/compactor.go

Lines 523 to 530 in 27d1c8a

	func (c *Compactor) compactUserWithRetries(ctx context.Context, userID string) error {
	var lastErr error
	retries := util.NewBackoff(ctx, util.BackoffConfig{
	MinBackoff: c.compactorCfg.retryMinBackoff,
	MaxBackoff: c.compactorCfg.retryMaxBackoff,
	MaxRetries: c.compactorCfg.CompactionRetries,
	})

An option may be:

1. Add BucketStores.syncUserBlocksWithRetries() similarly to the compactor one
2. Call syncUserBlocksWithRetries() like this:

func (u *BucketStores) SyncBlocks(ctx context.Context) error {
	return u.syncUsersBlocks(ctx, func(ctx context.Context, s *store.BucketStore) error {
		return s.syncUserBlocksWithRetries(ctx, s)
	})
}

[LeviHarrison]

@pracucci Hi! I'd like to take a shot at implementing this. I'm assuming I should just add those fields to the tsdb.BucketStoreConfig, or should they go somewhere else?

[pracucci]

@pracucci Hi! I'd like to take a shot at implementing this. I'm assuming I should just add those fields to the tsdb.BucketStoreConfig, or should they go somewhere else?

Thanks @LeviHarrison. I would start with an hardcoded backoff config like we did in the compactor (or other places too). 3 retries should be fine.