# Laboratory Data Storage Format Evaluation - Final Report

Generated by task 11 (report.py) from the measured benchmark matrix (task 09) and the extrapolation model (task 10). Every number at scales beyond the 150 MB corpus is a PROJECTION and is labeled as such; scaling assumptions are listed in section 6.

## 1. Executive summary

**Recommendation: PostgreSQL as the system of record.** The measured matrix and the projections confirm the expected outcome:

- **PostgreSQL** is the only format whose seven benchmark queries stay below the 1-hour practicality bound at every projected scale (600 GB, 1.2 TB, 6 TB); at 6 TB it needs one 16-core node with 551 GB RAM at an estimated $33,586 (bill of materials, prices as of 2026-07-11). Partitioned parallel scans and the track_summary materialized view are the scaling levers.
- **CSV** remains the canonical raw archive: byte-reproducible, instrument-native, and the best compressor (tar.gz ~2.8x); its full scans exceed 1 h beyond ~600 GB, so it is an archive, not a query layer.
- **SQLite** is an excellent single-user store up to ~600 GB (all queries OK), then its single-threaded scans degrade (Q5/Q7 IMPRACTICAL at 1.2 TB).
- **Hybrid JSON-LD** (metadata + sourceFile links, 1.8 MiB at 150 MB corpus) is the exchange format; the FULL variant re-reads the whole corpus per scan and becomes impractical past 600 GB.
- **RDF (materialized triplestore)** fails at scale: 11.7x storage blow-up, 5 of 7 queries FAIL at 6 TB, and the hot set needs 13.7 TB RAM across 14 nodes (~$527,732 projected). Semantic access should be a virtual layer (e.g. Ontop OBDA) over PostgreSQL instead.

## 2. Methodology snapshot

35 cells (Q1-Q7 x 5 formats); per cell 1 warm-up + 3 measured runs in isolated subprocesses sampled with psutil at 50 ms; randomized cell order; 30-min timeout; every run validated against canonical results (PostgreSQL cross-validated with SQLite, tolerance 1e-9). Medians are reported; raw runs and min/max live in out/bench/results_raw.csv. The cold-cache pass is skipped on this Windows host and marked as such. PostgreSQL runs on a remote host; its wall times include the LAN round-trip, and server-side metrics come from pg_stat_statements and Prometheus (docs/rules/bench-methodology.md section 6).

## 3. Environment

```
benchmark_host_os: Windows-2025Server-10.0.26100-SP0
benchmark_host_cpu: Intel64 Family 6 Model 141 Stepping 1, GenuineIntel
benchmark_host_cores_logical: 4
benchmark_host_cores_physical: 4
benchmark_host_ram_gib: 8.0
python: 3.12.10
benchmark_host_disks: VMware Virtual NVMe Disk / SSD / 100 GB
sqlite: 3.49.1
psycopg: 3.3.4
pyoxigraph: 0.5.9
ijson: 3.5.1
psutil: 7.2.2
postgresql_server: PostgreSQL 16.14 (Ubuntu 16.14-0ubuntu0.24.04.1) on x86_64-pc-linux-gnu, compiled by gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0, 64-bit
postgresql_host: 192.168.10.73 (remote; 4 cores / 3.8 GiB, see .done/06.ok)
cold_cache_pass: skipped (Windows host, no page-cache drop - bench-methodology section 6)
pg_backend_sampling: not possible for a remote server; Prometheus windows + pg_stat_statements deltas instead
```

## 4. Storage footprint

| format | measured | vs CSV | archival (compressed) | projected @600 GB | projected @1.2 TB | projected @6 TB |
|---|---|---|---|---|---|---|
| CSV | 149 MB | 1.00x | 53 MB | 600 GB | 1.20 TB | 6.00 TB |
| JSON-LD | 311 MB | 2.09x | 57 MB | 1.26 TB | 2.51 TB | 13 TB |
| SQLite | 123 MB | 0.83x | 55 MB | 498 GB | 996 GB | 4.98 TB |
| PostgreSQL | 397 MB | 2.67x | 53 MB | 1.60 TB | 3.21 TB | 16 TB |
| RDF triplestore | 1.74 GB | 11.70x | 126 MB | 7.02 TB | 14 TB | 70 TB |

![C1 - disk footprint](charts/c1_disk_footprint.png)

## 5. Measured performance (150 MB corpus)

| format | query | wall time | peak RSS | CPU util | read |
|---|---|---|---|---|---|
| CSV | q1 | 0.06 s | 4 MB | 0% | 0.0 MB |
| CSV | q2 | 0.17 s | 20 MB | 16% | 2.5 MB |
| CSV | q3 | 1.15 s | 21 MB | 24% | 31.8 MB |
| CSV | q4 | 0.71 s | 21 MB | 22% | 16.8 MB |
| CSV | q5 | 1.09 s | 21 MB | 23% | 31.8 MB |
| CSV | q6 | 0.28 s | 20 MB | 17% | 1.2 MB |
| CSV | q7 | 1.27 s | 21 MB | 23% | 32.5 MB |
| JSON-LD | q1 | 0.11 s | 22 MB | 10% | 1.4 MB |
| JSON-LD | q2 | 1.48 s | 24 MB | 24% | 53.1 MB |
| JSON-LD | q3 | 9.94 s | 24 MB | 25% | 297.6 MB |
| JSON-LD | q4 | 9.92 s | 24 MB | 25% | 297.8 MB |
| JSON-LD | q5 | 9.96 s | 24 MB | 25% | 297.6 MB |
| JSON-LD | q6 | 9.79 s | 24 MB | 24% | 297.8 MB |
| JSON-LD | q7 | 10 s | 24 MB | 24% | 297.9 MB |
| SQLite | q1 | 0.06 s | 4 MB | 0% | 0.0 MB |
| SQLite | q2 | 0.06 s | 4 MB | 0% | 0.0 MB |
| SQLite | q3 | 0.11 s | 4 MB | 0% | 0.0 MB |
| SQLite | q4 | 0.06 s | 4 MB | 0% | 0.0 MB |
| SQLite | q5 | 0.77 s | 25 MB | 23% | 117.9 MB |
| SQLite | q6 | 0.06 s | 4 MB | 0% | 0.0 MB |
| SQLite | q7 | 0.71 s | 25 MB | 23% | 93.0 MB |
| PostgreSQL | q1 | 0.22 s | 42 MB | 18% | 4.1 MB |
| PostgreSQL | q2 | 0.22 s | 42 MB | 20% | 4.1 MB |
| PostgreSQL | q3 | 0.22 s | 42 MB | 18% | 4.1 MB |
| PostgreSQL | q4 | 0.22 s | 42 MB | 18% | 4.1 MB |
| PostgreSQL | q5 | 0.44 s | 42 MB | 10% | 4.1 MB |
| PostgreSQL | q6 | 0.22 s | 42 MB | 19% | 4.1 MB |
| PostgreSQL | q7 | 0.33 s | 42 MB | 12% | 4.1 MB |
| RDF triplestore | q1 | 0.11 s | 32 MB | 17% | 3.6 MB |
| RDF triplestore | q2 | 2.81 s | 85 MB | 24% | 375.6 MB |
| RDF triplestore | q3 | 46 s | 264 MB | 25% | 5445.1 MB |
| RDF triplestore | q4 | 20 s | 168 MB | 25% | 2694.6 MB |
| RDF triplestore | q5 | 60 s | 264 MB | 25% | 5390.5 MB |
| RDF triplestore | q6 | 0.17 s | 46 MB | 16% | 15.9 MB |
| RDF triplestore | q7 | 41 s | 264 MB | 25% | 5380.5 MB |

Sub-interval note: cells faster than the 50 ms sampling cadence (fast SQLite/CSV point reads) under-report client RSS/CPU; wall times are exact.

![C2 - RAM requirement](charts/c2_ram_requirement.png)

Per-query charts: [Q1](charts/c3_q1_wall_time.png), [Q2](charts/c3_q2_wall_time.png), [Q3](charts/c3_q3_wall_time.png), [Q4](charts/c3_q4_wall_time.png), [Q5](charts/c3_q5_wall_time.png), [Q6](charts/c3_q6_wall_time.png), [Q7](charts/c3_q7_wall_time.png)

## 6. Projections (600 GB / 1.2 TB / 6 TB)

All values in this section are projected, never measured. Laws per access pattern (spec 10): flat / index-depth log / linear result set / linear x depth / full scan / parallel partitioned scan; constants calibrated on the measured point after subtracting the per-format harness floor. Flags: IMPRACTICAL > 1 h, FAIL > 24 h.

Projected wall times at 600 GB:

| format | query | projected wall time | flag |
|---|---|---|---|
| CSV | q1 | 0.06 s | OK |
| CSV | q2 | 7.1 min | OK |
| CSV | q3 | 73.5 min | IMPRACTICAL |
| CSV | q4 | 44.1 min | OK |
| CSV | q5 | 69.6 min | IMPRACTICAL |
| CSV | q6 | 14.5 min | OK |
| CSV | q7 | 81.6 min | IMPRACTICAL |
| JSON-LD | q1 | 0.11 s | OK |
| JSON-LD | q2 | 92.0 min | IMPRACTICAL |
| JSON-LD | q3 | 11.0 h | IMPRACTICAL |
| JSON-LD | q4 | 11.0 h | IMPRACTICAL |
| JSON-LD | q5 | 11.0 h | IMPRACTICAL |
| JSON-LD | q6 | 10.9 h | IMPRACTICAL |
| JSON-LD | q7 | 11.2 h | IMPRACTICAL |
| SQLite | q1 | 0.06 s | OK |
| SQLite | q2 | 1.27 s | OK |
| SQLite | q3 | 5.7 min | OK |
| SQLite | q4 | 0.06 s | OK |
| SQLite | q5 | 47.5 min | OK |
| SQLite | q6 | 5.18 s | OK |
| SQLite | q7 | 43.7 min | OK |
| PostgreSQL | q1 | 0.22 s | OK |
| PostgreSQL | q2 | 0.22 s | OK |
| PostgreSQL | q3 | 3.42 s | OK |
| PostgreSQL | q4 | 11 s | OK |
| PostgreSQL | q5 | 3.7 min | OK |
| PostgreSQL | q6 | 4.06 s | OK |
| PostgreSQL | q7 | 115 s | OK |
| RDF triplestore | q1 | 0.11 s | OK |
| RDF triplestore | q2 | 3.0 h | IMPRACTICAL |
| RDF triplestore | q3 | 2.1 d | FAIL |
| RDF triplestore | q4 | 22.1 h | IMPRACTICAL |
| RDF triplestore | q5 | 2.8 d | FAIL |
| RDF triplestore | q6 | 5.9 min | OK |
| RDF triplestore | q7 | 45.3 h | FAIL |

Degradation by data size: [point read](charts/c4_degradation_point_read.png), [indexed / join](charts/c4_degradation_indexed.png), [full scan](charts/c4_degradation_full_scan.png)

## 7. Hardware sizing and cost at 600 GB (projected)

Bill-of-materials costing with street prices as of **2026-07-11** (sources below). Every configuration buys whole components: 64 GB DDR5 ECC RDIMM modules, 7.68 TB enterprise NVMe U.2 drives, and a priced 1U chassis; disk capacity includes 30% free-space headroom.

| format | configuration | chassis | RAM | disk | extra nodes | total |
|---|---|---|---|---|---|---|
| CSV | 1 x 1U Supermicro AS-1015CS-TNR (16 cores); 1 x 64 GB DDR5 ECC RDIMM; 1 x 7.68 TB NVMe U.2 | $4,618 | $1,887 | $3,995 | $0 | $10,500 |
| JSON-LD | 1 x 1U Supermicro AS-1015CS-TNR (16 cores); 1 x 64 GB DDR5 ECC RDIMM; 1 x 7.68 TB NVMe U.2 | $4,618 | $1,887 | $3,995 | $0 | $10,500 |
| SQLite | 1 x 1U Supermicro AS-1015CS-TNR (16 cores); 1 x 64 GB DDR5 ECC RDIMM; 1 x 7.68 TB NVMe U.2 | $4,618 | $1,887 | $3,995 | $0 | $10,500 |
| PostgreSQL | 1 x 1U Supermicro AS-1015CS-TNR (16 cores); 1 x 64 GB DDR5 ECC RDIMM; 1 x 7.68 TB NVMe U.2 | $4,618 | $1,887 | $3,995 | $0 | $10,500 |
| RDF triplestore | 2 x 1U Supermicro AS-1015CS-TNR (16 cores); 22 x 64 GB DDR5 ECC RDIMM; 2 x 7.68 TB NVMe U.2 | $4,618 | $41,514 | $7,990 | $4,618 | $58,740 |

Component prices (retrieved 2026-07-11):

- 64 GB DDR5-5600 ECC RDIMM (A-Tech, 2x32GB kit): $1,887 - <https://atechmemory.com/collections/ddr5-memory-ram>
- 7.68 TB NVMe U.2 1DWPD enterprise SSD (Dell 14-16G compatible, Cloud Ninjas): $3,995 - <https://cloudninjas.com/products/new-7-68tb-nvme-u-2-1dwpd-sie-2-5-enterprise-solid-state-drive-for-14th-15th-16th-gen-dell>
- 1U AMD EPYC 9004/9005 chassis (Broadberry starting configurations): Supermicro AS-1015CS-TNR $4,618, Supermicro AS-1115CS-TNR $6,272 - <https://www.broadberry.com/amd-epyc-9004-supermicro-servers>
- Market context - server DRAM is in a documented 2025-2026 price surge: <https://www.techpowerup.com/342331/server-dram-pricing-jumps-50-only-70-of-orders-getting-filled>

Full sizing at 1.2 TB and 6 TB: out/bench/hardware_sizing.csv. RDF exceeds the 1 TB single-node RAM ceiling already at 600 GB and reaches 14 nodes at 6 TB.

![C6 - cost vs performance](charts/c6_cost_vs_performance.png)

## 8. Weighted scoring

Subscores are 10 x best/value per metric; weights: search x5, RAM economy x2, disk economy x1; max 80. Search uses the inverse geometric mean of Q1-Q7 (FAIL at 600 GB zeroes the search subscore). Left: measured scale; right: projected 600 GB.

Measured scale (150 MB):

| format | search (x5) | RAM (x2) | disk (x1) | total (max 80) |
|---|---|---|---|---|
| SQLite | 50.0 | 16.9 | 10.0 | 76.9 |
| CSV | 15.5 | 20.0 | 8.3 | 43.8 |
| PostgreSQL | 26.2 | 10.0 | 3.1 | 39.3 |
| JSON-LD | 1.7 | 17.2 | 4.0 | 22.9 |
| RDF triplestore | 1.3 | 1.6 | 0.7 | 3.6 |

Projected at 600 GB:

| format | search (x5) | RAM (x2) | disk (x1) | total (max 80) |
|---|---|---|---|---|
| PostgreSQL | 50.0 | 5.5 | 3.1 | 58.6 |
| SQLite | 21.8 | 15.2 | 10.0 | 47.0 |
| CSV | 0.6 | 20.0 | 8.3 | 28.9 |
| JSON-LD | 0.1 | 20.0 | 4.0 | 24.0 |
| RDF triplestore | 0.0 | 0.2 | 0.7 | 0.9 |

![C5 - weighted score](charts/c5_weighted_score.png)

## 9. Use-case mapping

| Use case | Recommended format(s) |
|---|---|
| Interactive analysis (joins, aggregations) | PostgreSQL; SQLite acceptable single-user up to ~600 GB |
| Report generation (repeated summaries) | PostgreSQL (track_summary materialized view) |
| Search / filtering | PostgreSQL or SQLite (indexed); flat formats need full scans |
| Archiving | CSV tree + tar.gz (canonical raw) plus pg_dump of the system of record |
| Inter-lab exchange | Hybrid JSON-LD (metadata + sourceFile links to CSV) |
| Semantic / ontology queries | Virtual RDF layer over PostgreSQL (e.g. Ontop OBDA), not a materialized triplestore |

## 10. Limitations

- Single-node measurements on one Windows host + one remote PostgreSQL host; no cluster or cloud variance.
- The corpus is simulated (physically plausible models, fixed seed 20260711); real instrument data may have different value distributions, though volumes and shapes match the lab.
- Extrapolation uses analytic laws calibrated on a single 150 MB point; no intermediate-scale validation runs were performed.
- Client wall times include a per-format harness floor (interpreter start, imports, connection); the floor is assumed constant across scales. The cheapest cell of each format therefore projects flat.
- The cold-cache matrix pass is skipped (no page-cache drop on Windows); all measured runs are warm-cache.
- The pandas CSV variant was not measured; its RAM is O(n) and it is infeasible above roughly 10 GB (spec 10).
- RDF numbers reflect the compact modeling and a client that derives run-in / steady-state from raw cycles; SPARQL-side aggregation engines could shift (not remove) the scan penalty.
- Hardware prices are spot street prices retrieved 2026-07-11 during a documented server-DRAM price surge; re-cost by editing out/config/hw_prices.yaml and re-running tasks 10-11.

## Appendix A - process flow diagrams (task 02)

### D1 - Coupon assembly

```mermaid
flowchart LR
    A([Ti-6Al-4V Base 10x10x3 mm]) --> B[Cleaning]
    B --> B1([Rinsed in deionized water])
    B1 --> B2([Sonicated in cleaning solution])
    B --> C[Smearing Adhesive]
    C --> C1([Cr adhesive coating])
    C --> D[Vapor Deposition System<br/>Kurt J. Lesker PVD 200]
    D --> D1([PtAu sputter coating, gradient])
    D --> E([Test Coupon])
```

### D2 - Characterization and batch assembly

```mermaid
flowchart LR
    TC([Test Coupon]) --> S[SIMTRA simulation<br/>composition / atom energies]
    S --> OP[Optical profilometry<br/>film thickness 0.3-1.1 um]
    OP --> TW[Assemble Test Wafer<br/>x49 coupons]
    TW --> TB[Assemble Test Batch<br/>x3 wafers]
    TB --> B721[B721<br/>Pt 150 W / Au 50 W]
    TB --> B722[B722<br/>Pt 100 W / Au 100 W]
    TB --> B723[B723<br/>Pt 150 W / Au 50 W]
    TB --> B724[B724<br/>Pt 50 W / Au 150 W]
```

### D3 - Testing tree

```mermaid
flowchart TD
    TC([Test Coupon]) --> F[Friction<br/>RAPID 6-probe tribometer]
    F --> FA([Lab Air -> COF dataset mu_normal])
    F --> FN([Dry N2 -> COF dataset mu_dry_nit])
    TC --> NI[Nanoindentation<br/>Bruker TI980]
    NI --> NIR([Hardness, reduced modulus<br/>25 indents per coupon])
    TC --> AF[AFM]
    AF --> AFR([Topography -> Ra / Rq])
    TC --> XR[micro-XRF<br/>Bruker M4 Tornado]
    XR --> XRR([Pt/Au wt% map, 20x20 grid])
```

### D4 - Tribometer session sequence

```mermaid
sequenceDiagram
    actor Op as Operator
    participant Prep as Sample Prep
    participant Plan as Excel Test Plan
    participant SW as Control Software
    participant TR as RAPID Tribometer

    Op->>Prep: Request test
    Op->>Prep: Get samples
    Op->>Prep: Load samples onto plates
    Op->>Prep: Load ball holders (3 counterfaces per holder)
    Op->>Plan: Create test plan<br/>(Sample Plate, Plate Location, Sample ID, Save Location,<br/>Folder Name, X/Y/Z Offset, Load, Iterations)
    Op->>TR: Transfer plates to tribometer
    Op->>TR: Set up platter
    Op->>SW: Activate software
    Op->>SW: Load plan
    SW->>TR: Activate tribometer
    TR-->>SW: Run reciprocating tests (execution loop, see D5)
    Op->>SW: Software stop
    SW-->>Op: Save avg files
    Op->>TR: Remove plates and platters
    Op->>SW: Close software
    Op->>TR: Pull equipment out
```

### D5 - Execution loop

```mermaid
flowchart TD
    L[Load all counterfaces] --> P{5 plates}
    P -->|for each plate| PR{6 probes in parallel}
    PR -->|for each probe| C{4 coupons on coupon square}
    C -->|for each coupon| T[Draw track]
    T --> RC[Rotate counterface - fresh ball per track]
    RC -->|3 tracks per coupon| T
    T --> DONE([120 coupons x 3 tracks = 360 tracks per run])
```

### D6 - Data hierarchy

```mermaid
erDiagram
    DEPOSITION_RUN ||--|| SIMTRA_PROFILE : produces
    DEPOSITION_RUN ||--|| BATCH : creates
    BATCH ||--|{ WAFER : contains
    WAFER ||--|{ COUPON : contains
    COUPON ||--|| XRF_MAP : has
    COUPON ||--|| NANOINDENTATION : has
    COUPON ||--|| AFM : has
    COUPON ||--|| PROFILOMETRY : has
    COUPON ||--o{ TRACK : "friction coupons only"
    TRACK ||--|{ CYCLE : has
    TRACK ||--|{ LOOP_POINT : has
    TRACK ||--|| WEAR : has
```