How much revenue can a few minutes of search downtime cost during a flash sale?
Host swaps often break indexing pipelines, create stale facets, and make routing black holes that kill conversions. This problem hits small teams juggling CI/CD and inventory sync especially hard.
Operators need a low-risk, automatable cutover that keeps relevancy. They must preserve incremental updates and allow instant rollback. Run the checklist during one maintenance window.
To move an e-commerce search (Algolia or Solr) during a host swap without visible customer downtime, run a staged reindex. Create parallel indexes, stream incremental changes, run mirrored query traffic in canary, validate relevancy and latency, then switch DNS or load balancer with a scripted rollback. The reproducible playbook gives exact CLI and API snippets, an automated canary switch, rollback scripts, concrete validation queries, and pass/fail metrics. It shows how to move ecommerce search systems during host swaps.
Summary of process
Run this checklist to finish a zero-downtime search cutover in one maintenance window.
This summary gives ordered tasks and expected timing so the release manager can act fast. Each step is scriptable and has pass or fail criteria. Total swap time typically runs between 10 and 60 minutes depending on index size and mirror percentage.
Follow the numbered steps below in order. The canary observation window should be 5–10 minutes. Lower DNS TTL at T-72h before the swap.
Steps
- Provision parallel writable and copy schema and settings. Aim to keep an operational RPO of ≤1 second for live writes during the cutover window. Prepare and validate the parallel index at least 48–72 hours before the planned swap so dual-write, CDC, checksum parity and canary rehearsals complete before the maintenance window.
- Bulk copy then enable dual-write or CDC. Verify object counts and checksums. Time: full copy may take 10–120 minutes.
- Mirror 5–10% production queries to the target for 5–10 minutes. Validate top-k match and latency. Goal: top-5 match ≥99%.
- Perform atomic swap using Algolia moveIndex or Solr alias swap. Ramp traffic to 100% in 10% steps. Monitor for 30 minutes.
What the summary assumes
The site can dual-write or has an available CDC stream to capture deltas while bulk copying. The index schema is stable and the application can route queries through an alias or a reverse proxy for mirroring.
Prepare target index
Create a parallel index or collection and make it writable before cutover. Provision compute and storage in the target region (us-east-1 recommended for US traffic). Match shard and replica counts to the source. The goal is identical query routing and similar latency characteristics.
Export and copy schema, ranking rules, synonyms and facet definitions before sending any user traffic to the new cluster. Small differences in field types or tokenizers change relevance a lot.
Algolia setup details
Initialize a new index, copy settings, synonyms and replicas, then run a resumable bulk copy. Use the Admin API key for settings and the safe write key for objects.
Code snippets:
javascript
// copy settings and synonyms
const algoliasearch = require('algoliasearch');
const client = algoliasearch('ALG_APP_ID', 'ALG_ADMIN_KEY');
const src = client.initIndex('products_v1');
const dst = client.initIndex('products_v2');
await dst.setSettings(await src.getSettings());
await dst.saveSynonyms(await src.getSynonyms(), {replaceExistingSynonyms:true});
// resumable browse & upload
await src.browseObjects({batch:1000}, async hit => {
// accumulate and save in batches to dst
});
Solr setup details
Create a new collection and upload the same configset. Verify field types and analysis chains match exactly. Tune softCommit frequency for dual-write latency.
Commands:
bash
curl "http://new-solr:8983/solr/admin/collections?action=CREATE&name=products_v2&numShards=4&replicationFactor=2"
Give concrete CDC and dual-write wiring examples so incremental reindexing keeps parity during the swap. For transactional catalogs, stream DB changes into Kafka with Debezium. Consume into a transformer that normalizes payloads and sends batched updates to the target search engine. For Algolia use batch saveObjects and partialUpdateObjects to avoid replacing full records. For Solr push atomic updates to the update handler with softCommit tuning to balance visibility and fsync cost.
Example: stream events keyed by SKU to a transformer that emits JSON batches of 1,000. POST to /1/indexes/INDEX_NAME/batch for Algolia or to /solr/collection/update?softCommit=true for Solr with softCommit intervals tuned to match your SLA.
Track per-document lastWrite timestamps and run a small reconciliation job that samples 1,000 random SKUs to compare checksums and lastModified. This incremental reindex plus CDC approach keeps RPO low during bulk copy index operations. It ensures near real-time consistency without blocking queries.
One clear test can cover core queries and facets.
Run mirrored canary tests
Mirror a small percentage (5–10%) of production search traffic and validate relevance and latency before switching all users. Collect metrics for an observation window of 5–10 minutes and require steady-state metrics before proceeding.
Use a representative query corpus of 50–200 weighted queries that includes high-value SKUs, common facets and typo cases. Weight queries by revenue and facet usage.
How to mirror traffic
Nginx simple mirror configuration for non-impactful mirroring:
nginx
location /search {
proxy_pass http://old_search;
mirror /mirror;
}
location = /mirror {
internal;
proxy_pass http://new_search;
proxy_set_header X-Canary "true";
}
Envoy offers percentage-based request mirroring via the route's request_mirror_policy for production-grade mirroring.
Validation queries and metrics
Collect these metrics per query and in aggregate: top-k match rate (top-5), precision at k delta, p50 and p95 latencies, and error rate delta. Define thresholds before testing. Require top-5 match ≥99%, p95 latency delta ≤50ms, and error delta ≤0.1%.
A repeatable post-swap validation playbook closes the loop on canary traffic mirroring and A/B checks. Capture a weighted control corpus of 200–1,000 queries with revenue, facet and typo tags. Use it to run three automated suites: mirrored top-k agreement checks, latency and error regression tests, and small A/B or canary cohorts for live business signals.
Automate pass or fail rules. For example, require top-5 match ≥99% and p95 delta ≤50ms for two consecutive 5-minute windows before any further ramp. Implement these checks as CI jobs or a lightweight service that queries both backends. Compare results and emit a single boolean decision to drive the traffic ramp or an immediate rollback.
Storing the query corpus and results makes post-mortem analysis and relevance tuning deterministic.
Execute an atomic alias or index move once canary metrics meet thresholds and remain stable for the observation window. For Algolia use moveIndex to atomically replace the live index. For Solr update the collection alias to point to the new collection. Ramp traffic slowly and monitor business metrics.
Automate rollback triggers that revert the alias or index move when defined thresholds are breached during or after ramp.
Algolia atomic swap
Algolia provides an atomic index move API that swaps index names server side without downtime.
Snippet:
javascript
// move products_v2 into products atomically
await client.moveIndex('products_v2','products');
If moveIndex fails or metrics exceed thresholds, move the prior index back or restore from the saved alias snapshot.
Solr alias swap and verification
Create or overwrite an alias to change which collection serves queries atomically.
Commands:
bash
curl "http://solr:8983/solr/admin/collections?action=CREATEALIAS&name=products_alias&collections=products_v1"
curl "http://solr:8983/solr/admin/collections?action=CREATEALIAS&name=products_alias&collections=products_v2&overwrite=true"
Verify alias use with a health query and then increase traffic in controlled steps.
Automated rollback scripts are essential for low-risk cutovers. Provide a small, reproducible rollback routine that watches the same health metrics used for the canary. Use top-k agreement, p95 latency and error rate as triggers. On threshold breach, programmatically revert routing and index names.
For example, run a CI job or Lambda that checks metrics and calls Algolia's moveIndex to restore the old index. Or call Solr's Collections API to recreate the previous alias mapping. Include safeguards: run the revert only if the pre-swap index checksum and alias mapping files are intact and timestamped. Pair the automation with a short live audit that samples 500 recent queries and verifies the restored backend returns the pre-swap top-k ordering.
Including reversible CLI and API snippets in the rollback routine and preserving a read-only snapshot of the prior alias state reduces human error. This lets teams run a fully automatic search rollback strategy during high-stakes flash sales or promotions.
A quick audit finds parity issues fast.
Map relevance between Algolia and Solr
Align ranking rules, typo handling, synonyms and boosting to keep user-facing relevance during the swap. Small differences in typo tolerance, proximity boosts and field weightings can change conversion metrics by double digits on critical queries. Map each rule explicitly and test on the control queries.
Run a relevance parity check using top-k agreement and CTR proxy metrics. Require business acceptance, for example no more than 1% drop in checkout add-rate on canary users.
Key mapping pairs
- Typo handling: Algolia typo tolerance ↔ Solr edismax fuzziness and mm settings.
- Custom ranking: Algolia customRanking ↔ Solr function queries and boost parameters (bf and bq).
Synonyms, facets and normalization
Export synonyms from Algolia via the synonyms API and import into Solr's SynonymFilterFactory. Align facet field types and mincount to avoid missing buckets.
| Criterion |
Algolia |
Solr |
Action |
| Typo tolerance |
Automatic multi-typo settings |
edismax mm and fuzzy params |
Tune edismax mm and fuzziness to match typo rate |
| Custom ranking |
customRanking attributes |
bf and bq function queries |
Translate boosts to function queries |
| Synonyms |
Synonyms API |
SynonymFilterFactory |
Export, import and normalize casing |
Expect relevance mapping mistakes to change CTR by up to double digits for top queries when left unchecked. Always validate with control queries and a revenue-weighted sample.
The evidence shows attention to ranking parity reduces user-impact risk during swaps. An example case: a retailer mapped only field weights but skipped proximity boosts and saw a 12% drop in add-to-cart for phrase queries during a cutover. That issue is easy to detect with a top-k agreement test and prevent with matched phrase boosts.
DNS, CDN and geo-consistency
Sequence DNS TTL changes, CDN invalidation and geo-replica promotions to avoid split-brain and inconsistent responses. Lower DNS TTL to 60 seconds 48–72 hours before the swap. Pre-warm CDN edges and plan targeted invalidations instead of global purges to avoid cache stampedes.
Promote or warm geo-replicas in the target region only after local parity checks. Track which edge nodes serve old versus new backends during the observation window.
Timeline and concrete timings
- T-72h: lower DNS TTL to 60s and confirm propagation across Route 53 or the provider. This reduces DNS caching delays.
- T-48h: create and bulk-copy the target index and start dual-write or CDC.
- T-24h: begin mirror at 1–5% to warm caches and measure stability.
- T-0: perform atomic swap and run the 5–10 minute canary ramp.
Refer to AWS Route 53 docs for TTL handling and propagation nuances: AWS Route 53 Developer Guide.
CDN cache warming and targeted
Warm edges by replaying the high-frequency query set against the target index to populate edge caches. Invalidate only search result keys that include index references or surrogate keys to avoid broad purges.
Cloudflare and other CDN providers document edge caching patterns. Plan invalidation windows and stagger invalidations regionally if traffic is global. See a CDN primer: Cloudflare CDN learning.
Cutover flow
Provision target
Create index/collection, copy schema
Bulk copy
Bulk upload, checksum, dual-write
Mirror traffic
5–10% canary, validate top-k and latency
Atomic swap
moveIndex or alias swap, ramp to 100%
Observe & rollback
Auto-rollback triggers and incident logging
Frequently asked questions
How to avoid Algolia downtime during migration?
Use a parallel writable index, dual-write or CDC, and Algolia's moveIndex for atomic cutover. Mirror 5–10% production queries for 5–10 minutes and require top-5 agreement ≥99% before moving the index.
How is Algolia different from Solr during host swaps?
Algolia is managed and has atomic moveIndex and built-in replicas. Solr needs collection aliases and explicit replica promotion. Expect different operational costs and latency tradeoffs when choosing managed versus self-hosted.
What exact checks prove an index is synced?
Check object counts, sample-hash checksums, last update timestamps, and per-shard commit times. A practical parity check is sample-hash mismatch ≤0.1% across 1,000 random records.
What are reliable rollback triggers?
Rollback if top-5 match falls below 98% for more than 60 seconds. Rollback if p95 latency increases by more than 200ms for over 60 seconds. Rollback if error rate delta exceeds 0.5% for two consecutive 30-second buckets.
How much does Algolia index migration cost?
Cost varies by vendor plan, record counts, and feature use. Algolia pricing depends on records, operations and plan tier rather than a flat per-100k-query rate. Self-hosted Solr costs depend on instance size, replication and ops time. Estimate additional query volume during mirroring and use vendor billing calculators or a cloud TCO estimate for Solr.
Next steps and references
Run the steps above in a staging rehearsal first and collect the control queries and monitoring dashboards you will use during production. Keep an incident playbook with commands for alias swap reversal and DNS rollback.
Links and references used in this playbook: Algolia docs for moveIndex and synonyms: Algolia Documentation. Apache Solr official docs: Apache Solr. AWS Route 53 DNS guidance: AWS Route 53 Guide.
A rehearsal run of this playbook typically takes between 2 and 4 hours to complete for mid-sized catalogs. Plan for 24–72 hours of TTL and cache preparation before production cutover.
Call the SRE or Release manager immediately if the canary shows relevance regressions greater than the defined business threshold during ramp.
Authoritativeness signals and compliance
- PCI-DSS v4.Guidance issued affects how cardholder data is handled in search logs and must be considered when logging queries.
- The GDPR is in force and affects how personal data in search records is processed and migrated.
- Kubernetes reached 1.Zero initially and many modern Solr deployments run on Kubernetes.
