Data centres, storage and backup

In short

A data centre is a purpose-built facility that houses servers, storage and network gear with the power, cooling, cabling, fire protection and physical security they need to run continuously. Storage design decides how data is laid out across disks for capacity, speed and fault tolerance (RAID, SAN/NAS/object), while backup and disaster recovery decide how much data and time an organisation can afford to lose (RPO and RTO). The topic appears in university papers on computer networks, system administration and cloud computing, and is asked in almost every infrastructure, cloud, NOC and IT-support interview — TCS, Infosys, Wipro, HCLTech, Cognizant, LTIMindtree, and data-centre operators such as CtrlS, Yotta, NTT-Netmagic, STT GDC India, Sify, Nxtra by Airtel and AdaniConneX. In India it is also a compliance subject: the DPDP Act, 2023, CERT-In's 2022 directions and RBI's payment-data localisation rule all dictate where data sits and how long logs are kept.

Explained

Data centre: types, tiers and standards

• Types by ownership: *on-premises* (enterprise's own room), *colocation* (rack/cage rented in a shared facility, customer owns the hardware), *hyperscale* (cloud providers' own mega-facilities), *edge* (small facilities near users for low latency), *managed/hosted*.
• Uptime Institute Tier classification (facility topology, not a quality grade):
  • Tier I – single path for power and cooling, no redundancy; ≈99.671% availability (~28.8 h downtime/yr).
  • Tier II – redundant capacity components (N+1), single distribution path; ≈99.741% (~22 h).
  • Tier III – concurrently maintainable: any component can be taken out for service without stopping IT; ≈99.982% (~1.6 h).
  • Tier IV – fault tolerant: 2N, any single failure (including a distribution path) is absorbed automatically; ≈99.995% (~26 min).
• TIA-942 is the telecom infrastructure standard for data centres: spaces (entrance room, MDA/HDA/EDA, zone distribution), cabling, pathways, rated 1–4. EN 50600 is the European equivalent. ASHRAE TC 9.9 gives thermal classes (A1–A4) with a recommended inlet range of about 18–27 °C and 20–80% RH for most IT equipment.
• Indian codes and practice: National Building Code of India, 2016 (structure, Part 4 fire and life safety), seismic design per IS 1893, earthing per IS 3043, lightning protection per the IS/IEC 62305 series, electrical safety under the CEA Safety Regulations, energy performance under the Energy Conservation Act with BEE/ECBC provisions. Management systems: ISO/IEC 27001 (information security), ISO 22301 (business continuity), ISO 50001 (energy). Specific IS numbers are revised often — quote the subject if the edition is uncertain.
• Physical layout: 42U/47U racks (1U = 44.45 mm, 19-inch mounting width), raised floor or overhead containment, hot-aisle/cold-aisle arrangement with blanking panels, cable trays separated for power and data, water-leak and temperature sensors, two-factor access control, mantrap, CCTV with retention.

Power and cooling

• Power chain: utility feeders (ideally from two substations) → transformer → ATS → DG sets for mains failure → UPS for the seconds in between and for conditioning → PDU (floor/rack) or busway → rack power strips → dual-corded servers fed from A and B paths.
• UPS types: *offline/standby* (cheap, home use), *line-interactive* (small offices), online double-conversion (data centres — zero transfer time, full conditioning). Batteries: VRLA traditionally, lithium-ion increasingly (smaller, longer life, higher capex).
• Redundancy notation: N = just enough capacity; N+1 = one spare unit; 2N = two complete independent systems; 2N+1 = two systems plus a spare. Tier III typically needs N+1 with dual paths; Tier IV needs 2N.
• Cooling: heat removed equals IT power consumed. CRAC (direct-expansion) and CRAH (chilled water) units, in-row and rear-door coolers, chillers with cooling towers, *free cooling* (air- or water-side economiser) where ambient allows, liquid/direct-to-chip and immersion cooling for high-density AI racks (30–100+ kW per rack).
• Airflow discipline: cold-aisle or hot-aisle containment, no bypass or recirculation, sealed floor cut-outs; raising the supply temperature within ASHRAE limits saves chiller energy.
• Fire protection: very early smoke detection (aspirating, e.g. VESDA), clean-agent suppression (inert gas, FK-5-1-12/HFC blends) rather than water in IT halls, pre-action sprinklers outside, compartmentation, EPO.
• Efficiency metrics: PUE (total facility energy ÷ IT energy), DCiE, WUE (water), CUE (carbon). Indian hyperscale facilities commonly report design PUE around 1.3–1.5; legacy server rooms are often 2.0+.

Storage: architectures and media

• DAS (direct-attached): disks inside or cabled to one server (SATA/SAS/NVMe). Cheapest, lowest latency, but capacity is stranded in that server.
• NAS (network-attached): a file server reached over Ethernet/IP using NFS (Unix/Linux) or SMB/CIFS (Windows). Shares folders; easy to manage; good for user data and home directories.
• SAN (storage area network): presents block devices (LUNs) over a dedicated fabric — Fibre Channel (8/16/32/64 Gb, HBAs, FC switches, zoning), iSCSI (SCSI over TCP/IP, cheap), FCoE, and modern NVMe-oF (NVMe over Fabrics — FC, RoCE or TCP) for microsecond latency. Used for databases, virtualisation and clustering.
• Object storage: data stored as objects with metadata in a flat namespace, accessed by HTTP REST APIs — Amazon S3, Azure Blob, Google Cloud Storage, MinIO, Ceph RGW. Infinitely scalable, cheap, ideal for backups, media, logs, analytics; not suitable for a transactional database.
• Block vs file vs object: block = raw volumes the OS formats; file = a hierarchy with POSIX/SMB semantics; object = immutable whole objects with versioning and lifecycle rules.
• Media: HDD (7.2k/10k/15k rpm, high capacity, ~75–200 IOPS each), SSD/NVMe flash (tens of thousands of IOPS, sub-millisecond latency), LTO tape for archive (LTO-9 is 18 TB native, ~45 TB compressed; LTO-10 is newer), optical/WORM for regulatory archives.
• Features: thin provisioning, auto-tiering (hot data on flash, cold on HDD/object), compression, deduplication (inline or post-process), snapshots (point-in-time, copy-on-write or redirect-on-write), clones, replication (synchronous = zero data loss but distance-limited by latency; asynchronous = some lag, any distance), software-defined storage (Ceph, VSAN, Gluster) and hyperconverged infrastructure (compute + storage on the same commodity nodes).
• Performance vocabulary: IOPS (operations/s, matters for databases and VDI), throughput (MB/s, matters for backup and video), latency (ms, matters for user experience), queue depth, block size, read/write mix, cache hit ratio.

RAID and erasure coding

• RAID 0 (striping) – speed, no redundancy; any disk failure loses all data.
• RAID 1 (mirroring) – two copies; 50% usable; survives one disk per mirror.
• RAID 5 – striping with one parity block; usable = (N−1) disks; survives one failure; long rebuilds on large drives create a second-failure risk.
• RAID 6 – dual parity; usable = (N−2); survives two failures; preferred for drives above ~4 TB.
• RAID 10 (1+0) – mirrored pairs then striped; 50% usable; best random-write performance and fast rebuild.
• Write penalty (back-end I/Os per host write): RAID 1/10 = 2, RAID 5 = 4 (read data, read parity, write data, write parity), RAID 6 = 6.
• Hot spare, degraded mode, rebuild time; RAID protects against *disk failure only* — not deletion, corruption, ransomware or fire. RAID is not a backup.
• Erasure coding (k data + m parity fragments spread across nodes/sites) gives RAID-6-like protection with far less overhead at scale — e.g. 10+4 costs 1.4× raw instead of 2–3× for replication, at the price of CPU and rebuild network traffic.

Backup, archive and disaster recovery

• Backup types: full (everything), incremental (changes since the last backup of any type — small window, long restore chain), differential (changes since the last full — medium window, two-step restore), synthetic full (server merges a full with incrementals), forever-incremental with periodic merge, CDP (continuous data protection, journalled writes).
• Retention: GFS (grandfather-father-son) — daily, weekly, monthly, yearly copies; legal/statutory retention drives the yearly tail.
• 3-2-1 rule: 3 copies, on 2 different media, 1 copy off-site. Extended 3-2-1-1-0: one copy immutable or air-gapped, and zero errors in verification. Immutability (object lock/WORM) is the main defence against ransomware encrypting the backups too.
• RPO (recovery point objective) = maximum tolerable data loss, in time; RTO (recovery time objective) = maximum tolerable downtime. They drive the design: tape = hours to days; async replication = minutes; sync replication + clustering = near zero.
• DR site types: *cold* (space and power only), *warm* (hardware ready, data restored on demand), *hot* (fully running, data replicated), *active-active* (both sites serve traffic). Cloud DR and DRaaS reduce the idle capex.
• Backup media and targets: disk-to-disk-to-tape (D2D2T), purpose-built dedupe appliances, cloud cold tiers (Glacier/Archive — cheap storage, retrieval fee and delay), tape vault off-site.
• Operational essentials: backup windows that do not clash with production load, *restore* testing (a backup never test-restored is unverified), catalogue/metadata protection, encryption at rest and in flight, documented runbooks, periodic DR drills — the RBI expects regulated entities to test DR and the business continuity plan regularly.
• Indian regulatory context:
  • Digital Personal Data Protection Act, 2023 — purpose limitation, security safeguards, breach notification to the Data Protection Board, erasure when the purpose ends; the Act permits transfers abroad except to countries the Central Government restricts. Implementation is phased under the 2025 Rules, so check current notifications.
  • CERT-In Directions, 2022 — report specified cyber incidents within 6 hours, maintain ICT logs for 180 days within India, synchronise clocks to NIC/NPL NTP.
  • RBI (2018) — all payment system data must be stored only in India (processing abroad allowed with data brought back); separate RBI/SEBI/IRDAI norms on outsourcing, cloud and BCP-DR.
  • Government workloads use MeitY-empanelled cloud service providers; Indian capacity is concentrated in Mumbai and Navi Mumbai (cable landings), Chennai, Hyderabad, Bengaluru, Pune, Noida and Kolkata; national live capacity is in the low gigawatts and growing fast, so quote current figures only from a dated source.

Key points

• Tier III vs Tier IV – concurrently maintainable (service without downtime) vs fault tolerant (survives any single failure automatically).
• PUE – total facility energy ÷ IT energy; 1.0 is ideal, below ~1.5 is good practice, 2.0+ means half the power is overhead.
• N+1 / 2N – one spare unit vs two fully independent systems; Tier IV needs 2N and dual active paths.
• Online double-conversion UPS – the only UPS class used for IT halls; DG handles sustained outages, UPS the gap.
• Hot-aisle/cold-aisle with containment – stops hot and cold air mixing; blanking panels are mandatory discipline.
• 1 TR = 3.517 kW – cooling load in refrigeration tons equals IT heat load in kW ÷ 3.517.
• DAS/NAS/SAN/object – local disk / file shares over IP / block LUNs over a fabric / REST-addressed objects.
• iSCSI vs Fibre Channel – IP on existing Ethernet and cheap, vs a dedicated lossless fabric with lower latency and zoning.
• RAID usable capacity – 0: N·C; 1 and 10: N·C/2; 5: (N−1)C; 6: (N−2)C.
• Write penalty – 2 for RAID 1/10, 4 for RAID 5, 6 for RAID 6; it decides how many disks a workload really needs.
• RAID ≠ backup – it survives disk failure, not deletion, corruption, ransomware or site loss.
• Erasure coding – k+m fragments; cheaper than replication at petabyte scale, costlier in CPU and rebuild traffic.
• Full / incremental / differential – backup window shrinks in that order; restore complexity grows.
• RPO vs RTO – how much data you may lose vs how long you may be down; they are independent targets.
• 3-2-1-1-0 – three copies, two media, one off-site, one immutable/air-gapped, zero verification errors.
• Synchronous vs asynchronous replication – zero RPO but distance-limited by round-trip latency, vs any distance with a lag.
• CERT-In – 6-hour incident reporting and 180-day log retention inside India.
• DPDP Act, 2023 + RBI 2018 – security safeguards and breach reporting for personal data; payment system data stored in India.

Formulas and examples

Example 1 — PUE and the cost of inefficiency. A Mumbai colocation hall draws 680 kW total with a 400 kW IT load; tariff ₹8/kWh.

• PUE = 680 ÷ 400 = 1.70; DCiE = 1/1.70 = 58.8%.
• Non-IT (overhead) load = 280 kW → annual overhead energy = 280 × 8,760 = 2,452,800 kWh → cost = ₹19,622,400 ≈ ₹1.96 crore/yr.
• Retrofitting containment and raising supply temperature brings PUE to 1.40 → total 560 kW, overhead 160 kW; saving = 120 kW × 8,760 = 1,051,200 kWh ≈ ₹84.1 lakh/yr.
• Cooling to be designed for the IT heat: 400 ÷ 3.517 = 113.7 TR, plus margin and N+1 units.

Example 2 — RAID capacity and the write penalty. Eight 4 TB SAS disks, each sustaining 150 IOPS; workload is 3,000 host IOPS at 70% read / 30% write.

• Usable capacity: RAID 0 = 32 TB; RAID 10 = 16 TB; RAID 5 = 7 × 4 = 28 TB; RAID 6 = 6 × 4 = 24 TB.
• Reads = 2,100 IOPS, writes = 900 IOPS.
• RAID 5 back-end = 2,100 + (900 × 4) = 5,700 IOPS → 5,700 ÷ 150 = 38 disks needed.
• RAID 10 back-end = 2,100 + (900 × 2) = 3,900 IOPS → 26 disks.
• So RAID 5 saves capacity but needs 12 more spindles for the same performance — hence flash or RAID 10 for write-heavy databases.

Example 3 — availability of a power path. UPS A = 99.95%, cooling = 99.90%, network = 99.99%, all three needed (series).

• A = 0.9995 × 0.9990 × 0.9999 = 0.99840 → 99.840%.
• Downtime = (1 − 0.99840) × 8,760 = 0.00160 × 8,760 ≈ 14.0 h/yr.
• Now duplicate the UPS in parallel (2N): A_UPS = 1 − (0.0005)² = 0.99999975 → overall A = 0.99999975 × 0.9990 × 0.9999 ≈ 0.99890 → downtime ≈ 9.6 h/yr. Cooling is now the weak link, so it must be made N+1 too — redundancy only helps where the failure actually is.

Example 4 — backup window and repository sizing. 20 TB production data, 3% daily change, weekly full + 6 daily incrementals, 4 weeks retained, 10 Gbps backup LAN at 70% efficiency, dedupe+compression 5:1, 20% headroom.

• Effective throughput = 10 × 0.70 = 7 Gbps = 875 MB/s.
• Full backup = 20 TB = 20 × 1,048,576 MB = 20,971,520 MB ÷ 875 = 23,967 s ≈ 6 h 40 min — fits an 8-hour window.
• Daily incremental = 0.03 × 20 TB = 0.6 TB = 629,146 MB ÷ 875 = 719 s ≈ 12 min.
• Logical data retained = 4 fulls (80 TB) + 24 incrementals (24 × 0.6 = 14.4 TB) = 94.4 TB.
• After 5:1 reduction = 18.88 TB; with 20% headroom = ≈ 22.7 TB of repository.

Example 5 — RPO, data at risk and replication bandwidth. The same system replicates asynchronously to a Chennai DR site every 15 minutes.

• Change rate = 0.6 TB/day = 614.4 GB/day → 25.6 GB/h.
• Data at risk at a 15-minute RPO = 25.6 × 0.25 = 6.4 GB (one quarter-hour's writes).
• Steady-state bandwidth = 614.4 × 1024 × 8 ÷ 86,400 = 5,033,164.8 ÷ 86,400 ≈ 58.3 Mbps; provision ~2× (≈120 Mbps) for peaks and resynchronisation.
• If the business demands RPO = 0, synchronous replication is required — and the Mumbai–Chennai round-trip latency (tens of milliseconds) would make it unusable, so a metro DR site is needed instead.

Where it is used

• Colocation and hyperscale operators: CtrlS (Hyderabad), Yotta (Navi Mumbai), STT GDC India, NTT-Netmagic, Sify, Nxtra by Airtel, AdaniConneX, Web Werks and Equinix — roles in facility operations, electrical/HVAC engineering, DCIM and NOC.
• IT services and cloud migration: TCS, Infosys, Wipro, HCLTech, Cognizant, Capgemini and LTIMindtree staff storage administration, backup administration (Veeam, Commvault, NetBackup, Rubrik), and lift-and-shift to AWS/Azure/GCP Mumbai and Hyderabad regions.
• Banking, NBFC and insurance: primary DC plus near-DR and far-DR, RBI-mandated BCP and DR drills, payment data localised in India, core-banking storage on all-flash SAN.
• Telecom, OTT and e-commerce: CDN edge caches, object storage for media libraries, petabyte-scale analytics lakes.
• Government and public sector: State Data Centres, NIC/NICSI facilities, MeitY-empanelled cloud for e-governance, CERT-In log retention within India.
• Job roles: data centre technician/facility engineer, storage administrator, backup and DR engineer, cloud infrastructure engineer, SRE, IT auditor (ISO 27001/22301).

Interview questions

• Q: What is the difference between Tier III and Tier IV? A: Tier III is concurrently maintainable — any component can be serviced without stopping IT — while Tier IV is fault tolerant, surviving any single unplanned failure automatically with 2N power and cooling paths.
• Q: What does PUE of 1.5 mean? A: For every 1 kW consumed by IT equipment, the facility draws 1.5 kW in total — 0.5 kW goes to cooling, UPS losses, lighting and ancillaries.
• Q: Difference between NAS and SAN? A: NAS serves files over IP using NFS or SMB and the storage device owns the filesystem; a SAN serves raw block LUNs over FC/iSCSI/NVMe-oF and the host formats them, which is why databases and hypervisors prefer SAN.
• Q: Why is RAID 6 preferred over RAID 5 for large drives? A: Rebuilding a multi-terabyte drive takes many hours during which a second failure would destroy the array; RAID 6's dual parity survives that second failure.
• Q: What is the RAID 5 write penalty? A: Each host write becomes four back-end I/Os — read old data, read old parity, write new data, write new parity — so write-heavy workloads need far more spindles.
• Q: Is RAID a backup? A: No. RAID protects against disk hardware failure only; it faithfully replicates deletions, corruption, ransomware encryption and is lost entirely if the site is lost.
• Q: Difference between incremental and differential backup? A: An incremental copies changes since the previous backup of any type (small, but restore needs the whole chain); a differential copies changes since the last full (bigger each day, but restore needs only the full plus the latest differential).
• Q: Explain RPO and RTO with an example. A: RPO is tolerable data loss — a 15-minute RPO means up to 15 minutes of transactions may be lost; RTO is tolerable downtime — a 4-hour RTO means service must be back within four hours. Nightly tape may give RTO 8 h and RPO 24 h.
• Q: When would you use synchronous replication, and what limits it? A: When RPO must be zero, for example core banking; it is limited by distance because every write waits for the remote acknowledgement, so round-trip latency caps it at metro distances.
• Q: What is the 3-2-1-1-0 rule? A: Three copies of data, on two different media, one off-site, one immutable or air-gapped, with zero errors on verification/restore testing.
• Q: How do snapshots differ from backups? A: A snapshot is a point-in-time view on the same array, instant and space-efficient but dependent on the primary storage; a backup is an independent copy on separate media that survives loss of that array.
• Q: What are CERT-In's key obligations for an Indian data centre? A: Report specified cyber incidents within 6 hours of noticing them, retain ICT logs for 180 days within India, and keep clocks synchronised to NIC/NPL NTP servers.
• Q: Why is object storage used for backups rather than a database? A: Objects are written and read whole over HTTP with versioning, lifecycle and object-lock immutability — perfect for large write-once data, but the latency and lack of in-place partial updates make it unsuitable for transactional workloads.

Common mistakes

• Treating RAID, snapshots or replication as a backup — replication copies ransomware and deletions to the DR site within seconds.
• Quoting RAID capacity correctly but ignoring the write penalty, so the array meets the TB target and misses the IOPS target.
• Confusing RPO with RTO, or assuming a small RPO automatically gives a small RTO.
• Computing backup time from the link's nominal bit rate without converting bits to bytes or allowing for protocol and disk overhead.
• Designing redundancy on only one element (two UPS units but a single chiller or a single utility feeder) and still claiming Tier IV.
• Never test-restoring: an untested backup, an unverified catalogue or an unrehearsed DR plan is an assumption, not a control.
• Forgetting blanking panels, floor-cutout sealing and containment, then blaming hot spots on insufficient CRAC capacity.
• Ignoring Indian localisation and retention rules when choosing a cloud region — payment system data must stay in India and logs must be retained for 180 days in-country.

Related topics

Network administration and troubleshooting · Windows Server and Active Directory · Cloud computing and virtualisation · Operating systems: file systems and I/O · Information security and cryptography · Database administration and high availability