ezharjan commited on
Commit
25fa86b
·
verified ·
1 Parent(s): d5fee4e

Add DOI 10.57967/hf/10841

Browse files
Files changed (3) hide show
  1. CITATION.cff +31 -30
  2. README.md +629 -627
  3. scripts/package_huggingface.py +1 -0
CITATION.cff CHANGED
@@ -1,30 +1,31 @@
1
- cff-version: 1.2.0
2
- message: "If you use this dataset, please cite it as below. The toolkit that produced it has its own citation in the source repository."
3
- type: dataset
4
- title: "LocalNet-Dynamics: A Longitudinal Local Area Network Telemetry Dataset"
5
- authors:
6
- - family-names: "Ezharjan"
7
- given-names: "Alexander"
8
- version: 1.0.0
9
- date-released: "2026-10-09"
10
- license: MIT
11
- url: "https://huggingface.co/datasets/ezharjan/LocalNetDynamics"
12
- repository-code: "https://github.com/Ezharjan/LocalNetDynamicsDataset"
13
- abstract: >-
14
- A longitudinal telemetry dataset of a residential local area network,
15
- sampled every 30 seconds around the clock from a wired desktop for 19 days.
16
- Each sample records ICMP latency, jitter and loss to a reference host and to
17
- the gateway, TCP connect time, cached and recursive DNS response time, HTTP
18
- time-to-first-response and hourly WAN throughput tests, with time-of-day
19
- features and rule-based reference labels for forecasting, anomaly detection
20
- and quality-of-service research.
21
- keywords:
22
- - network telemetry
23
- - latency
24
- - jitter
25
- - packet loss
26
- - DNS
27
- - throughput
28
- - quality of service
29
- - anomaly detection
30
- - time series
 
 
1
+ cff-version: 1.2.0
2
+ message: "If you use this dataset, please cite it as below. The toolkit that produced it has its own citation in the source repository."
3
+ type: dataset
4
+ title: "LocalNet-Dynamics: A Longitudinal Local Area Network Telemetry Dataset"
5
+ authors:
6
+ - family-names: "Ezharjan"
7
+ given-names: "Alexander"
8
+ version: 1.0.0
9
+ date-released: "2026-10-09"
10
+ license: MIT
11
+ doi: 10.57967/hf/10841
12
+ url: "https://huggingface.co/datasets/ezharjan/LocalNetDynamics"
13
+ repository-code: "https://github.com/Ezharjan/LocalNetDynamicsDataset"
14
+ abstract: >-
15
+ A longitudinal telemetry dataset of a residential local area network,
16
+ sampled every 30 seconds around the clock from a wired desktop for 19 days.
17
+ Each sample records ICMP latency, jitter and loss to a reference host and to
18
+ the gateway, TCP connect time, cached and recursive DNS response time, HTTP
19
+ time-to-first-response and hourly WAN throughput tests, with time-of-day
20
+ features and rule-based reference labels for forecasting, anomaly detection
21
+ and quality-of-service research.
22
+ keywords:
23
+ - network telemetry
24
+ - latency
25
+ - jitter
26
+ - packet loss
27
+ - DNS
28
+ - throughput
29
+ - quality of service
30
+ - anomaly detection
31
+ - time series
README.md CHANGED
@@ -1,627 +1,629 @@
1
- ---
2
- pretty_name: LocalNet-Dynamics
3
- license: mit
4
- language:
5
- - en
6
- tags:
7
- - tabular
8
- - timeseries
9
- - networking
10
- - telemetry
11
- - latency
12
- - jitter
13
- - packet-loss
14
- - dns
15
- - wifi
16
- - throughput
17
- - qos
18
- - anomaly-detection
19
- task_categories:
20
- - time-series-forecasting
21
- - tabular-classification
22
- - tabular-regression
23
- annotations_creators:
24
- - machine-generated
25
- source_datasets:
26
- - original
27
- size_categories:
28
- - 10K<n<100K
29
- configs:
30
- - config_name: default
31
- data_files:
32
- - split: train
33
- path: dataset/localnet_dynamics.parquet
34
- ---
35
-
36
- # LocalNet-Dynamics
37
-
38
- A longitudinal, high-resolution telemetry dataset of a residential local area
39
- network, sampled every 30 seconds around the clock from a wired desktop for
40
- 19 days (2026-09-20 to 2026-10-09). Each of the 54,617 rows is one sampling
41
- epoch and combines ICMP latency, loss and jitter to a reference host and to
42
- the gateway, TCP connect time, cached and recursive DNS response time, HTTP
43
- time-to-first-response and hourly WAN throughput tests with time-of-day
44
- features and rule-based reference labels.
45
-
46
- - Author and point of contact: Alexander Ezharjan
47
- - Dataset: https://huggingface.co/datasets/ezharjan/LocalNetDynamics
48
- - Source code (collector, build pipeline, analysis): https://github.com/Ezharjan/LocalNetDynamicsDataset
49
- - License: MIT (code and data)
50
- - Version: 1.0.0
51
-
52
- ## Intended uses
53
-
54
- - Time-series forecasting of latency, jitter and loss. The capture shows
55
- strong persistence (lag-1 autocorrelation of the log round-trip time 0.81,
56
- still 0.42 after three hours) and almost no 24-hour periodicity (0.09 at a
57
- one-day lag), so it rewards models that track regime changes rather than
58
- seasonal ones.
59
- - Unsupervised and semi-supervised anomaly detection, evaluated against the
60
- reference labels and the episode tables in `dataset/analysis.json`.
61
- - QoS classification (for example, whether a sample meets the requirements of
62
- real-time voice, cloud gaming or video conferencing).
63
- - Root-cause studies: the gateway probe, the TCP probe and the
64
- cached-versus-recursive DNS timings make it possible to separate first-hop
65
- degradation from path changes and from resolver degradation. The capture
66
- contains worked examples of each.
67
- - Studies of the relationship between ICMP loss and application-visible
68
- latency (HTTP retransmission modes, DNS tails).
69
-
70
- ## Collection site
71
-
72
- | Item | Value |
73
- |---|---|
74
- | Vantage point | one desktop PC, host label `desktop-eth` |
75
- | Link between client and router | wired Ethernet (the collector could not read the media type, see the limitations) |
76
- | Client operating system | Windows 11 (build 26200), Python 3.12.12 |
77
- | Gateway and resolver | the home router, a private address on the `192.168.1.0/24` LAN |
78
- | Time zone | US Eastern, UTC-4 for the whole run |
79
- | Reference host | `8.8.8.8` (ICMP and TCP port 443) |
80
- | Access technology, advertised plan, router model | not recorded |
81
- | Household usage during the run | normal use, not annotated |
82
-
83
- ## Files
84
-
85
- | File | Content |
86
- |---|---|
87
- | `dataset/localnet_dynamics.parquet` | consolidated dataset, one row per sample, typed columns (the file served by the dataset viewer) |
88
- | `dataset/localnet_dynamics.csv` | the same table as CSV |
89
- | `dataset/summary.json` | row counts, coverage, per-metric statistics, per-run label thresholds |
90
- | `dataset/analysis.json` | every number quoted in this card: profiles, regimes, episodes, conditional statistics, autocorrelation, speed-test quality |
91
- | `dataset/raw/localnet_desktop-eth_20260920_123407.csv` | raw measurements of the collection run, exactly as written by the collector |
92
- | `dataset/raw/localnet_desktop-eth_20260920_123407.json` | configuration, environment, detected network and running summary of that run |
93
- | `figures/png/`, `figures/pdf/` | the ten figures below, as 200 dpi PNG and as vector PDF |
94
- | `scripts/` | collector, builder, annotation tool, analysis, plotting and packaging scripts |
95
- | `tests/` | pytest suite |
96
-
97
- The consolidated files are produced from the raw files by
98
- `scripts/build_dataset.py`; the raw files are never edited. No manual event
99
- annotations were recorded during this run, so there is no `events.csv`.
100
-
101
- ## Findings
102
-
103
- The numbers in this section are computed by `scripts/analyze.py` and stored in
104
- `dataset/analysis.json`.
105
-
106
- ### Overview
107
-
108
- ![Overview timeline](figures/png/overview_timeline.png)
109
-
110
- 54,617 samples from 2026-09-20 12:34 to 2026-10-09 15:49 local time, 99.1 % of
111
- the samples expected at the 30-second interval. 345 gaps are exactly 60 s,
112
- the tick skipped after an hourly speed test; 78 gaps are longer, up to 122 s,
113
- 53 of them after a speed test that overran two ticks and the rest after
114
- samples whose probes waited for their timeouts. The median round-trip time
115
- to 8.8.8.8 is 11.1 ms (95th percentile 22.3 ms), the median first-hop
116
- round-trip time 0.2 ms, the median TCP handshake 14.5 ms.
117
-
118
- ### Two regimes
119
-
120
- A single least-squares change point in the daily share of samples with ICMP
121
- loss falls on 2026-09-30 and divides the capture into two ten-day halves:
122
-
123
- | | Sep 20 to Sep 29 | Sep 30 to Oct 9 |
124
- |---|---|---|
125
- | samples | 27,020 | 27,597 |
126
- | samples with at least one lost echo | 23.2 % | 5.3 % |
127
- | mean packet loss | 2.98 % | 0.57 % |
128
- | round-trip time, median / p95 / p99 | 11.9 / 24.3 / 28.4 ms | 10.8 / 14.2 / 25.5 ms |
129
- | jitter, median / p95 | 2.22 / 9.22 ms | 0.78 / 4.56 ms |
130
- | HTTP response, median / p95 | 87.7 / 1,094 ms | 77.6 / 157 ms |
131
- | HTTP responses slower than 1 s | 5.87 % | 1.09 % |
132
- | recursive DNS, median | 62.5 ms | 56.2 ms |
133
- | samples with any label | 38.9 % | 9.2 % |
134
-
135
- ![Regimes](figures/png/regimes.png)
136
-
137
- ![Labels](figures/png/labels.png)
138
-
139
- ### Loss: daytime, weekdays, upstream of the gateway
140
-
141
- ![Diurnal profile](figures/png/diurnal_profile.png)
142
-
143
- ![Calendar heat maps](figures/png/calendar_heatmaps.png)
144
-
145
- 14.1 % of all samples lost at least one of their ten echoes; 6,309 lost
146
- exactly one, 1,092 two, 230 three, 41 four. On weekdays the hourly share of
147
- samples with loss rises from 6 to 8 % between midnight and 05:00 to 25.2 % at
148
- 08:00 and stays above 20 % until 16:00; on weekends it stays between 6 % and
149
- 14 % all day. Weekdays as a whole have 15.2 % of samples with loss, weekends
150
- 10.5 %. Outside the four outages, the gateway lost a packet in exactly one
151
- sample, so the loss sits beyond the first hop.
152
-
153
- ![Loss anatomy](figures/png/loss_anatomy.png)
154
-
155
- Loss is only weakly bursty: a sample that follows a lossy sample is lossy with
156
- probability 0.30 against a base rate of 0.14, and the autocorrelation of the
157
- loss indicator is 0.18 at lag 1. The other probes show what a lost echo means:
158
-
159
- | | no loss | 1 of 10 lost | 2 or more lost |
160
- |---|---|---|---|
161
- | samples | 46,907 | 6,309 | 1,383 |
162
- | round-trip time, median | 11.0 ms | 11.8 ms | 12.6 ms |
163
- | jitter, median | 1.00 ms | 2.12 ms | 3.29 ms |
164
- | TCP handshake, median | 14.5 ms | 14.3 ms | 14.4 ms |
165
- | HTTP responses slower than 1 s | 2.54 % | 7.47 % | 16.34 % |
166
- | recursive DNS slower than 200 ms | 1.75 % | 5.22 % | 9.60 % |
167
-
168
- The HTTP response time has a smooth body with a median of 81.9 ms and a 95th
169
- percentile of 271 ms, then discrete modes: 2.48 % of responses fall between
170
- 1.0 and 1.4 s and 0.55 % between 3.9 and 4.4 s. Their timing matches the
171
- retransmission timers of the operating system's DNS client and TCP stack (one
172
- second, then multiples of it), and they occur three to six times more often in
173
- samples that also lost an echo, so they are lost packets waiting out a timer
174
- rather than slow servers: the mechanism by which an echo lost in a ten-second
175
- burst becomes a one-second stall for an application.
176
-
177
- ### Latency episodes: path changes and a first-hop event
178
-
179
- ![Degradation anatomy](figures/png/degradation_anatomy.png)
180
-
181
- Runs of the high-latency label, merged across gaps of up to ten samples, give
182
- nine episodes of 30 minutes or more; they contain 88.4 % of all high-latency
183
- labels. Each is classified from the probe that moved: *first hop* if the
184
- median gateway round-trip time exceeds the run's gateway threshold (5.2 ms),
185
- *path change* if the median minimum ICMP time or the median fastest TCP
186
- handshake rises at least 5 ms above its run median (10.0 and 12.3 ms), and
187
- *queueing* otherwise.
188
-
189
- | start (local) | end (local) | hours | ICMP min | ICMP mean | TCP fastest | gateway | kind |
190
- |---|---|---|---|---|---|---|---|
191
- | Mon Sep 21 22:34 | Tue Sep 22 01:05 | 2.52 | 20.0 | 26.4 | 25.7 | 0.0 | path change |
192
- | Tue Sep 22 01:27 | 02:41 | 1.23 | 10.0 | 20.4 | 20.1 | 0.0 | path change |
193
- | Tue Sep 22 02:47 | 03:44 | 0.95 | 11.0 | 21.1 | 19.5 | 0.0 | path change |
194
- | Thu Sep 24 23:51 | Fri Sep 25 01:16 | 1.42 | 11.0 | 22.3 | 20.1 | 0.0 | path change |
195
- | Sat Sep 26 14:07 | 14:54 | 0.78 | 14.0 | 23.5 | 18.1 | 11.4 | first hop |
196
- | Sat Sep 26 15:01 | 15:32 | 0.52 | 14.0 | 22.0 | 18.1 | 11.2 | first hop |
197
- | Sat Sep 26 16:36 | 18:17 | 1.68 | 14.0 | 23.7 | 20.7 | 10.6 | first hop |
198
- | Sun Sep 27 00:00 | 06:16 | 6.28 | 17.0 | 25.6 | 23.8 | 0.0 | path change |
199
- | Tue Sep 29 21:21 | Wed Sep 30 07:34 | 10.22 | 20.0 | 24.8 | 24.5 | 0.0 | path change |
200
-
201
- Values are medians over the episode, in milliseconds; an end time without a
202
- date is on the start date. The six path changes fall on four nights and all
203
- lie between 21:21 and 07:34. In every one of them the fastest TCP handshake
204
- sits 7 to 13 ms above its 12.3 ms baseline, the gateway is untouched, and the
205
- DNS probe, whose queries the gateway forwards to the upstream resolver, stays
206
- within 4 ms of its 36 ms median, so the extra delay lies on the route towards
207
- 8.8.8.8, beyond the first hop and beyond the point where the path to the
208
- resolver branches off. The three first-hop episodes are one event on the
209
- afternoon of Saturday September 26: the gateway answered in 10 to 11 ms
210
- instead of 0.2 ms, every other probe inherited the delay, and that day holds
211
- 507 of the 560 gateway-degraded labels.
212
-
213
- ### Outages
214
-
215
- | start (local) | duration | samples | gateway reachable |
216
- |---|---|---|---|
217
- | Mon Sep 21 13:03 | 2.0 min | 4 | no |
218
- | Thu Sep 24 07:10 | 2.5 min | 5 | no |
219
- | Sat Sep 26 13:11 | 2.5 min | 5 | no |
220
- | Tue Oct 6 02:51 | 2.0 min | 4 | no |
221
-
222
- Every outage took the gateway down together with the target, with name
223
- resolution failing as well (`getaddrinfo failed` in the `notes` column), so
224
- all four were first-hop events. The ISP path never failed on its own during
225
- the capture. Outages account for 18 samples, 0.03 % of the dataset.
226
-
227
- ### The TCP/ICMP split
228
-
229
- ![TCP versus ICMP](figures/png/tcp_vs_icmp.png)
230
-
231
- Over the whole capture the fastest TCP handshake exceeds the minimum ICMP time
232
- by 2.27 ms (median). From 22:00 on October 6 the share of samples in which it
233
- exceeds it by more than 5 ms jumps from 7.2 % to 63.7 %, and the fastest
234
- handshake becomes bimodal (first quartile 12.9 ms, median 18.6 ms, third
235
- quartile 22.2 ms) while the ICMP minimum stays at 10 ms. The ICMP and TCP
236
- probes address the same host, so the difference is in how the two kinds of
237
- flow are routed; a monitor that relies on ping alone would not see the change.
238
-
239
- ### DNS
240
-
241
- The "cached" probe asks the gateway resolver for `example.com`, which it has
242
- answered before, yet its median is 36.0 ms (5th percentile 17 ms; only 3.4 %
243
- of answers arrive within 15 ms). The gateway therefore forwards every query
244
- upstream rather than serving it from a local cache, and the column measures
245
- the upstream resolver's responsiveness. The recursive probe (a random name
246
- under `example.com`) has a median of 59.2 ms, 20.2 ms above the cached probe
247
- in the same sample; 2.35 % of recursive lookups take longer than 200 ms and
248
- the 99th percentile is 1,266 ms. 24 samples recorded a DNS timeout, 14 of
249
- them during the four outages (the other four outage samples failed with a
250
- socket error instead).
251
-
252
- ### Memory
253
-
254
- ![Autocorrelation](figures/png/autocorrelation.png)
255
-
256
- | lag | 30 s | 5 min | 30 min | 3 h | 12 h | 24 h |
257
- |---|---|---|---|---|---|---|
258
- | log round-trip time | 0.81 | 0.76 | 0.68 | 0.42 | 0.18 | 0.09 |
259
- | jitter | 0.44 | 0.42 | 0.36 | 0.25 | 0.23 | 0.17 |
260
- | lost-echo indicator | 0.18 | 0.17 | 0.16 | 0.14 | 0.04 | 0.10 |
261
- | fastest TCP handshake | 0.28 | 0.15 | 0.14 | 0.09 | 0.04 | 0.02 |
262
- | recursive DNS | 0.04 | 0.02 | 0.02 | 0.02 | 0.00 | 0.01 |
263
- | HTTP | 0.06 | 0.05 | 0.03 | 0.04 | 0.02 | 0.01 |
264
-
265
- Latency is persistent for hours and has no daily cycle at this site; DNS and
266
- HTTP timings are memoryless at every lag.
267
-
268
- ### WAN throughput
269
-
270
- ![Speed tests](figures/png/speedtest.png)
271
-
272
- 460 hourly speed tests ran through the `speedtest-cli` Python fallback with no
273
- pinned server and used 162 different servers. 79 tests could not measure the
274
- server's latency (the fallback encodes that failure as a value of 600,000 ms
275
- or more; the builder sets it to missing), 7 reported a download of 0 Mbit/s
276
- and 4 an upload of 0 Mbit/s. Download: median 34.7 Mbit/s, 90th percentile
277
- 183 Mbit/s, maximum 480.7 Mbit/s; among the ten most used servers the
278
- per-server median ranges from 20.9 to 371.8 Mbit/s. Upload: median
279
- 28.6 Mbit/s, 90th percentile 40.3 Mbit/s, maximum 54.2 Mbit/s. The median
280
- download is 61.5 Mbit/s for the 95 tests started between 00:00 and 04:59 and
281
- 22.0 Mbit/s for the 156 started between 09:00 and 16:59. The download column is a lower bound on
282
- the access link and depends heavily on the server; the upload column is more
283
- consistent. Loaded latency was not measured because the official Ookla CLI was
284
- not installed.
285
-
286
- ## Measurement protocol
287
-
288
- One sample is taken at every tick of a fixed schedule (`start + k * interval`).
289
- Within a sample the following probes run concurrently:
290
-
291
- - ICMP: `ping_count` (10) echo requests to the target and `gateway_ping_count`
292
- (5) to the default gateway, one second apart, each waiting at most
293
- `ping_timeout` (2 s) for its reply. Per-reply round-trip times are parsed
294
- from the output of the operating system's `ping`.
295
- - TCP: `tcp_count` (5) TCP handshakes to `target:443`, 0.2 s apart, timed with
296
- a high-resolution clock from `connect()` start to completion.
297
- - DNS: two UDP queries (type A, recursion desired) sent directly to the
298
- configured resolver, bypassing the operating-system cache. The first asks for
299
- `dns_host` (`example.com`), which the resolver serves from its cache after
300
- the first sample and therefore measures the resolver's responsiveness. The
301
- second asks for a random 12-hexadecimal-character subdomain of `dns_host`,
302
- which can never be cached and therefore measures a full recursive
303
- resolution. A response with rcode NOERROR or NXDOMAIN is a valid measurement.
304
- - HTTP: one GET request for `http_url`
305
- (`http://www.msftconnecttest.com/connecttest.txt`) with the system proxy
306
- bypassed, timed until the response headers have been received.
307
- - Wi-Fi: link metrics read from the operating system
308
- (`netsh wlan show interfaces`, `iw dev <if> link`, `system_profiler
309
- SPAirPortDataType`).
310
-
311
- After the concurrent probes, and only on their own cadence (every hour in the
312
- published run), the heavy tests run: an Ookla speed test and/or an `iperf3`
313
- test. Because they run after the latency probes of the same sample, they do
314
- not influence that sample's latency columns; a sample that includes a speed
315
- test usually overruns the interval, and the collector then skips to the next
316
- tick.
317
-
318
- The default gateway, the interface type and the resolver are detected at start
319
- and re-detected every 15 minutes (or within a minute after the gateway probe
320
- loses every packet). System sleep is inhibited for the duration of the run.
321
-
322
- Run configuration of the published capture: interval 30 s, 10 echoes to
323
- `8.8.8.8` and 5 to the gateway with a 2 s timeout, 5 TCP handshakes to port
324
- 443, DNS probes against the gateway resolver with a 3 s timeout, HTTP probe
325
- with a 5 s timeout, Wi-Fi metrics enabled, SSID not recorded, speed test every
326
- hour with no pinned server, no iperf3 server, collector version 2.0.0.
327
-
328
- ## Schema
329
-
330
- Empty cells mean "not measured or failed"; the `notes` column records the
331
- reason (for example `ping: no replies`, `dns recursive: timeout`,
332
- `speedtest (ookla) failed: ...`). Times are in milliseconds unless the column
333
- name says otherwise.
334
-
335
- ### Identity and time
336
-
337
- | Column | Type | Description |
338
- |---|---|---|
339
- | `run_id` | string | collection run identifier, `<host_label_>YYYYMMDD_HHMMSS` in local time of the run start |
340
- | `host_label` | string | free-text label of the vantage point (`--host-label`), empty if not given |
341
- | `sample_index` | int | 0-based row number within the run |
342
- | `timestamp_utc` | string | sample start, ISO 8601 UTC, e.g. `2026-09-05T14:00:00Z` |
343
- | `timestamp_local` | string | sample start in the collector's local time with UTC offset, e.g. `2026-09-05T10:00:00-04:00` |
344
- | `epoch_s` | int | sample start as Unix time (seconds) |
345
- | `elapsed_s` | float | seconds since the start of the run |
346
- | `link_type` | string | `wifi`, `ethernet`, `other` or `unknown`: the interface that carries the default route |
347
- | `gateway_ip` | string | default gateway (private address) |
348
- | `dns_server` | string | resolver used for the DNS probes |
349
- | `target` | string | host used for the ICMP and TCP probes |
350
-
351
- ### Run context (added by the builder)
352
-
353
- | Column | Type | Description |
354
- |---|---|---|
355
- | `collector_version` | string | version of `collect_data.py` that wrote the run |
356
- | `collector_os` | string | `Windows`, `Linux` or `macOS` |
357
- | `interval_s` | float | configured sampling interval of the run |
358
- | `rtt_resolution_ms` | float | resolution of the ICMP columns: 1.0 on Windows, 0.001 elsewhere |
359
-
360
- ### Time features (added by the builder, from `timestamp_local`)
361
-
362
- | Column | Type | Description |
363
- |---|---|---|
364
- | `date_local` | string | local calendar date `YYYY-MM-DD` |
365
- | `hour_local` | int | local hour 0-23 |
366
- | `weekday_local` | int | 0 = Monday ... 6 = Sunday |
367
- | `is_weekend` | int | 1 on Saturday and Sunday |
368
- | `gap_before_s` | float | seconds since the previous sample of the same run; larger than `interval_s` after a skipped tick or a sleep/outage gap |
369
-
370
- ### Target latency (ICMP)
371
-
372
- | Column | Type | Description |
373
- |---|---|---|
374
- | `rtt_min_ms` | float | minimum round-trip time of the received replies |
375
- | `rtt_avg_ms` | float | arithmetic mean |
376
- | `rtt_median_ms` | float | median |
377
- | `rtt_max_ms` | float | maximum |
378
- | `rtt_stddev_ms` | float | population standard deviation of the received replies (the `mdev` of Linux `ping`) |
379
- | `jitter_ms` | float | mean absolute difference between consecutive received replies (packet delay variation in the sense of RFC 3393, applied to round-trip times); empty when fewer than two replies were received |
380
- | `packets_sent` | int | echo requests sent (`ping_count`) |
381
- | `packets_received` | int | replies received |
382
- | `packet_loss_pct` | float | `100 * (sent - received) / sent` |
383
-
384
- ### Gateway latency (ICMP to the default gateway)
385
-
386
- | Column | Type | Description |
387
- |---|---|---|
388
- | `gateway_rtt_min_ms` | float | minimum round-trip time |
389
- | `gateway_rtt_avg_ms` | float | mean |
390
- | `gateway_rtt_max_ms` | float | maximum |
391
- | `gateway_jitter_ms` | float | mean absolute consecutive difference |
392
- | `gateway_packet_loss_pct` | float | loss in percent |
393
-
394
- ### TCP, DNS and HTTP
395
-
396
- | Column | Type | Description |
397
- |---|---|---|
398
- | `tcp_connect_min_ms` | float | fastest of the TCP handshakes to `target:443` |
399
- | `tcp_connect_median_ms` | float | median handshake time |
400
- | `dns_resolve_ms` | float | response time of the resolver for `dns_host` (cache-served after the first sample) |
401
- | `dns_recursive_ms` | float | response time for a random subdomain of `dns_host` (full recursive lookup) |
402
- | `http_ms` | float | time to the response headers of `http_url`, including the operating-system name lookup, the TCP handshake and the request |
403
- | `http_status` | int | HTTP status code of that response |
404
-
405
- ### Wi-Fi link (empty on wired links)
406
-
407
- | Column | Type | Description |
408
- |---|---|---|
409
- | `wifi_ssid` | string | network name; only recorded with `--record-ssid` |
410
- | `wifi_signal_pct` | int | link quality 0-100: the value reported by Windows; on Linux and macOS derived from RSSI with the Windows mapping `quality = 2 * (rssi + 100)` clipped to 0-100 |
411
- | `wifi_rssi_dbm` | float | received signal strength: as reported on Linux and macOS; on Windows derived from the quality with `rssi = quality / 2 - 100` |
412
- | `wifi_rx_mbps` | float | negotiated receive PHY rate (Mbit/s) |
413
- | `wifi_tx_mbps` | float | negotiated transmit PHY rate (Mbit/s); equals the receive rate on macOS, which reports one rate |
414
- | `wifi_channel` | int | channel number |
415
- | `wifi_band_ghz` | float | 2.4, 5 or 6 |
416
- | `wifi_phy` | string | standard in use, e.g. `802.11ax` |
417
-
418
- ### WAN throughput (rows in which a speed test ran)
419
-
420
- | Column | Type | Description |
421
- |---|---|---|
422
- | `download_mbps` | float | download throughput (Mbit/s) |
423
- | `upload_mbps` | float | upload throughput (Mbit/s) |
424
- | `speedtest_ping_ms` | float | idle latency to the speed-test server as reported by the tool; missing when the `speedtest-cli` fallback could not measure it |
425
- | `download_loaded_latency_ms` | float | interquartile-mean latency measured during the download (Ookla CLI only) |
426
- | `upload_loaded_latency_ms` | float | interquartile-mean latency measured during the upload (Ookla CLI only) |
427
- | `speedtest_server_id` | int | Ookla server id |
428
- | `speedtest_server` | string | server name and location |
429
-
430
- ### LAN throughput (rows in which an iperf3 test ran)
431
-
432
- | Column | Type | Description |
433
- |---|---|---|
434
- | `lan_throughput_mbps` | float | TCP throughput to the LAN iperf3 server (Mbit/s) |
435
- | `lan_retransmits` | int | TCP retransmissions during the test, when reported |
436
-
437
- ### Reference labels (added by the builder)
438
-
439
- | Column | Type | Rule |
440
- |---|---|---|
441
- | `label_outage` | int | `packets_received == 0` while `packets_sent > 0` |
442
- | `label_packet_loss` | int | `0 < packet_loss_pct < 100` |
443
- | `label_high_latency` | int | `rtt_avg_ms > T_rtt` |
444
- | `label_high_jitter` | int | `jitter_ms > T_jitter` |
445
- | `label_gateway_degraded` | int | `gateway_rtt_avg_ms > T_gateway` or `gateway_packet_loss_pct > 0` |
446
- | `label_anomaly` | int | any of the above |
447
-
448
- The thresholds are computed per run from that run's own distribution with a
449
- robust rule, so they adapt to the baseline of the site without being pulled by
450
- the outliers they are meant to find:
451
-
452
- ```text
453
- T = median + max(k * 1.4826 * MAD, absolute floor, relative floor * median)
454
- ```
455
-
456
- with `k = 5`, the median absolute deviation `MAD`, and the floors
457
- 10 ms / 50 % for `rtt_avg_ms`, 5 ms / 100 % for `jitter_ms` and
458
- 5 ms / 100 % for `gateway_rtt_avg_ms`. A threshold is not computed (and the
459
- label stays 0) for runs with fewer than 20 valid values. For the published run
460
- the thresholds are 21.1 ms (`rtt_avg_ms`), 6.0 ms (`jitter_ms`) and 5.2 ms
461
- (`gateway_rtt_avg_ms`); the values used for each run are stored in
462
- `dataset/summary.json`. These labels are derived from the measurements
463
- themselves; they are reference points for evaluating detectors, not
464
- independent ground truth.
465
-
466
- ### Manual annotations and notes
467
-
468
- | Column | Type | Description |
469
- |---|---|---|
470
- | `event_category` | string | categories of the manual events covering the sample time, joined with ` \| `; one of `load`, `outage`, `maintenance`, `environment`, `isp`, `device`, `other` |
471
- | `event_description` | string | the corresponding free-text descriptions |
472
- | `sample_duration_s` | float | wall-clock time the sample's probes took |
473
- | `notes` | string | probe errors and status messages, `; `-separated |
474
-
475
- ## Validation performed by the builder
476
-
477
- - Rows with a malformed `timestamp_utc` are dropped and duplicate
478
- `(run_id, timestamp_utc)` pairs are reduced to the first occurrence (none in
479
- the published run).
480
- - Loss percentages outside 0-100 are set to missing (none in the published
481
- run).
482
- - `speedtest_ping_ms` values of 600,000 ms or more are set to missing. The
483
- `speedtest-cli` fallback scores each failed latency attempt as 3,600 s and
484
- averages three attempts over six, so one failed attempt already yields
485
- 600,000 ms and three yield 1,800,000 ms. The raw file keeps the original
486
- values; 79 of the 460 speed-test rows are affected. Collector version 2.0.1
487
- no longer records such values.
488
-
489
- ## Statistics
490
-
491
- <!-- BEGIN GENERATED STATISTICS -->
492
- Generated by `scripts/build_dataset.py` on 2026-10-10T05:42:01Z from 1 collection run(s).
493
-
494
- | Item | Value |
495
- |---|---|
496
- | Rows | 54,617 |
497
- | Columns | 67 |
498
- | First sample (UTC) | 2026-09-20T16:34:22Z |
499
- | Last sample (UTC) | 2026-10-09T19:49:43Z |
500
- | Span | 19.14 days |
501
- | Sampling coverage | 99.1 % of the samples expected at the configured interval |
502
- | Manual event annotations | 0 |
503
-
504
- | Run | Vantage | Link | OS | Interval | Rows | Span | Coverage |
505
- |---|---|---|---|---|---|---|---|
506
- | desktop-eth_20260920_123407 | desktop-eth | unknown (54617) | Windows | 30 s | 54,617 | 459.26 h | 99.1 % |
507
-
508
- | Metric | n | Median | Mean | p95 | Max |
509
- |---|---|---|---|---|---|
510
- | rtt_avg_ms | 54,599 | 11.1 | 12.581 | 22.3 | 288.6 |
511
- | rtt_min_ms | 54,599 | 10.0 | 10.432 | 14.0 | 263.0 |
512
- | rtt_max_ms | 54,599 | 13.0 | 17.869 | 32.0 | 373.0 |
513
- | jitter_ms | 54,599 | 1.0 | 2.306 | 7.667 | 59.778 |
514
- | packet_loss_pct | 54,617 | 0.0 | 1.763 | 10.0 | 100.0 |
515
- | gateway_rtt_avg_ms | 54,599 | 0.2 | 0.321 | 0.8 | 89.4 |
516
- | gateway_packet_loss_pct | 54,617 | 0.0 | 0.033 | 0.0 | 100.0 |
517
- | tcp_connect_median_ms | 54,598 | 14.484 | 19.311 | 32.127 | 1524.463 |
518
- | dns_resolve_ms | 54,595 | 36.026 | 43.594 | 74.878 | 3252.214 |
519
- | dns_recursive_ms | 54,590 | 59.242 | 87.275 | 106.31 | 25645.041 |
520
- | http_ms | 54,565 | 81.911 | 154.745 | 271.36 | 13111.02 |
521
- | wifi_signal_pct | 4 | 85.0 | 85.75 | 88.4 | 89.0 |
522
- | wifi_rssi_dbm | 4 | -57.5 | -57.125 | -55.8 | -55.5 |
523
- | wifi_rx_mbps | 4 | 866.7 | 866.7 | 866.7 | 866.7 |
524
- | download_mbps | 460 | 34.725 | 68.463 | 318.275 | 480.72 |
525
- | upload_mbps | 460 | 28.625 | 26.121 | 41.142 | 54.17 |
526
- | speedtest_ping_ms | 381 | 52.058 | 93.665 | 196.381 | 1167.189 |
527
-
528
- | Reference label | Rows | Share |
529
- |---|---|---|
530
- | label_outage | 18 | 0.033 % |
531
- | label_packet_loss | 7,692 | 14.084 % |
532
- | label_high_latency | 3,067 | 5.615 % |
533
- | label_high_jitter | 5,074 | 9.29 % |
534
- | label_gateway_degraded | 560 | 1.025 % |
535
- | label_anomaly | 13,036 | 23.868 % |
536
-
537
- <!-- END GENERATED STATISTICS -->
538
-
539
- ## Known limitations and biases
540
-
541
- - Single vantage point, single run. Absolute values reflect one site's access
542
- technology, router, cabling and ISP routing. Relative variation, the
543
- structure of episodes and the probe-to-probe relationships transfer better
544
- than absolute baselines.
545
- - `link_type` is `unknown` for every row. The PowerShell route query returned
546
- nothing on this machine, the collector fell back to `route print`, and that
547
- fallback does not report the media type. The vantage point is a wired
548
- desktop (host label `desktop-eth`, median first-hop round-trip time 0.2 ms).
549
- Four rows carry Wi-Fi link metrics (channel 44 or 157, 5 GHz, 802.11ac,
550
- 866.7 Mbit/s): a wireless adapter on the same machine was connected at those
551
- moments. Their first-hop round-trip times (0.0 to 1.2 ms) and target
552
- latencies match the wired baseline, so the default route stayed on the
553
- wired link.
554
- - ICMP resolution on Windows is 1 ms and sub-millisecond replies are stored as
555
- 0; jitter and standard deviation computed from such values contain
556
- quantisation noise, and the gateway columns are mostly 0 for this wired
557
- site. Use `tcp_connect_*` for sub-millisecond latency, or collect on
558
- Linux/macOS. `rtt_resolution_ms` states the resolution of each run.
559
- - ICMP is handled differently from user traffic by some routers and hosts
560
- (rate limiting, de-prioritisation). The TCP connect and HTTP columns measure
561
- the transport and application paths that applications actually use, and the
562
- October 6 TCP/ICMP split shows that the two can be routed differently.
563
- - The 10-echo burst characterises latency over roughly ten seconds of each
564
- interval; events shorter than that between samples are not observed.
565
- - The DNS "cached" probe is cached at the upstream resolver, not at the
566
- gateway: at this site the gateway forwards every query, so the column
567
- measures the upstream resolver plus the forwarding hop.
568
- - The DNS "recursive" probe generates NXDOMAIN answers for random names; a
569
- resolver that applies aggressive negative caching or that intercepts
570
- NXDOMAIN (ISP search redirection) still answers, but its timing reflects that
571
- behaviour.
572
- - The HTTP probe includes the operating system's name lookup for the URL's
573
- host, which is normally cached; its discrete modes near 1.1, 1.3 and 4.1 s
574
- are retransmission timers waiting out lost packets, not propagation or
575
- queueing delay.
576
- - Speed tests were run with the `speedtest-cli` fallback, which uses the
577
- Python interpreter's own HTTP transfers, and without a pinned server. The
578
- download column is a lower bound that depends strongly on the chosen
579
- server; 79 tests have no latency, 7 report 0 Mbit/s download and 4 report
580
- 0 Mbit/s upload. Loaded latency is absent. Speed tests also load the
581
- household link for about half a minute, which is visible to other devices,
582
- and they leave a 60-second gap in the 30-second grid.
583
- - The two halves of the capture differ markedly (see the findings). Models
584
- trained on one half should be evaluated on the other.
585
- - No manual event annotations were recorded, so the `event_*` columns are
586
- empty and the causes of the episodes are inferred from the probes only.
587
- - Gaps: system sleep, reboots and collector restarts appear as gaps
588
- (`gap_before_s`) or as separate runs; ticks skipped after an overrun also
589
- appear as gaps. The published run has none longer than 122 s.
590
- - Clock: timestamps come from the system clock; `elapsed_s` uses a monotonic
591
- clock and is unaffected by clock adjustments.
592
-
593
- ## Privacy
594
-
595
- The dataset does not contain public IP addresses, hostnames of household
596
- devices, MAC addresses, Wi-Fi network names or per-device traffic. Gateway and
597
- resolver addresses are private LAN addresses. The run metadata records the
598
- operating system build, the Python version, the time zone and the private LAN
599
- address of the collector. Speed-test rows name the test server, which was not
600
- pinned and therefore only indicates the broad region of the site.
601
-
602
- ## Citation
603
-
604
- If you use the data:
605
-
606
- ```bibtex
607
- @dataset{ezharjan_localnetdynamics_data_2026,
608
- author = {Ezharjan, Alexander},
609
- title = {LocalNet-Dynamics: A Longitudinal Local Area Network Telemetry Dataset},
610
- year = {2026},
611
- version = {1.0.0},
612
- publisher = {Hugging Face},
613
- url = {https://huggingface.co/datasets/ezharjan/LocalNetDynamics}
614
- }
615
- ```
616
-
617
- If you use the collector, the build pipeline or the analysis code:
618
-
619
- ```bibtex
620
- @software{ezharjan_localnetdynamics_code_2026,
621
- author = {Ezharjan, Alexander},
622
- title = {LocalNet-Dynamics: collector, build pipeline and analysis toolkit for longitudinal home-network telemetry},
623
- year = {2026},
624
- version = {1.0.0},
625
- url = {https://github.com/Ezharjan/LocalNetDynamicsDataset}
626
- }
627
- ```
 
 
 
1
+ ---
2
+ pretty_name: LocalNet-Dynamics
3
+ license: mit
4
+ language:
5
+ - en
6
+ tags:
7
+ - tabular
8
+ - timeseries
9
+ - networking
10
+ - telemetry
11
+ - latency
12
+ - jitter
13
+ - packet-loss
14
+ - dns
15
+ - wifi
16
+ - throughput
17
+ - qos
18
+ - anomaly-detection
19
+ task_categories:
20
+ - time-series-forecasting
21
+ - tabular-classification
22
+ - tabular-regression
23
+ annotations_creators:
24
+ - machine-generated
25
+ source_datasets:
26
+ - original
27
+ size_categories:
28
+ - 10K<n<100K
29
+ configs:
30
+ - config_name: default
31
+ data_files:
32
+ - split: train
33
+ path: dataset/localnet_dynamics.parquet
34
+ ---
35
+
36
+ # LocalNet-Dynamics
37
+
38
+ A longitudinal, high-resolution telemetry dataset of a residential local area
39
+ network, sampled every 30 seconds around the clock from a wired desktop for
40
+ 19 days (2026-09-20 to 2026-10-09). Each of the 54,617 rows is one sampling
41
+ epoch and combines ICMP latency, loss and jitter to a reference host and to
42
+ the gateway, TCP connect time, cached and recursive DNS response time, HTTP
43
+ time-to-first-response and hourly WAN throughput tests with time-of-day
44
+ features and rule-based reference labels.
45
+
46
+ - Author and point of contact: Alexander Ezharjan
47
+ - Dataset: https://huggingface.co/datasets/ezharjan/LocalNetDynamics
48
+ - DOI: https://doi.org/10.57967/hf/10841
49
+ - Source code (collector, build pipeline, analysis): https://github.com/Ezharjan/LocalNetDynamicsDataset
50
+ - License: MIT (code and data)
51
+ - Version: 1.0.0
52
+
53
+ ## Intended uses
54
+
55
+ - Time-series forecasting of latency, jitter and loss. The capture shows
56
+ strong persistence (lag-1 autocorrelation of the log round-trip time 0.81,
57
+ still 0.42 after three hours) and almost no 24-hour periodicity (0.09 at a
58
+ one-day lag), so it rewards models that track regime changes rather than
59
+ seasonal ones.
60
+ - Unsupervised and semi-supervised anomaly detection, evaluated against the
61
+ reference labels and the episode tables in `dataset/analysis.json`.
62
+ - QoS classification (for example, whether a sample meets the requirements of
63
+ real-time voice, cloud gaming or video conferencing).
64
+ - Root-cause studies: the gateway probe, the TCP probe and the
65
+ cached-versus-recursive DNS timings make it possible to separate first-hop
66
+ degradation from path changes and from resolver degradation. The capture
67
+ contains worked examples of each.
68
+ - Studies of the relationship between ICMP loss and application-visible
69
+ latency (HTTP retransmission modes, DNS tails).
70
+
71
+ ## Collection site
72
+
73
+ | Item | Value |
74
+ |---|---|
75
+ | Vantage point | one desktop PC, host label `desktop-eth` |
76
+ | Link between client and router | wired Ethernet (the collector could not read the media type, see the limitations) |
77
+ | Client operating system | Windows 11 (build 26200), Python 3.12.12 |
78
+ | Gateway and resolver | the home router, a private address on the `192.168.1.0/24` LAN |
79
+ | Time zone | US Eastern, UTC-4 for the whole run |
80
+ | Reference host | `8.8.8.8` (ICMP and TCP port 443) |
81
+ | Access technology, advertised plan, router model | not recorded |
82
+ | Household usage during the run | normal use, not annotated |
83
+
84
+ ## Files
85
+
86
+ | File | Content |
87
+ |---|---|
88
+ | `dataset/localnet_dynamics.parquet` | consolidated dataset, one row per sample, typed columns (the file served by the dataset viewer) |
89
+ | `dataset/localnet_dynamics.csv` | the same table as CSV |
90
+ | `dataset/summary.json` | row counts, coverage, per-metric statistics, per-run label thresholds |
91
+ | `dataset/analysis.json` | every number quoted in this card: profiles, regimes, episodes, conditional statistics, autocorrelation, speed-test quality |
92
+ | `dataset/raw/localnet_desktop-eth_20260920_123407.csv` | raw measurements of the collection run, exactly as written by the collector |
93
+ | `dataset/raw/localnet_desktop-eth_20260920_123407.json` | configuration, environment, detected network and running summary of that run |
94
+ | `figures/png/`, `figures/pdf/` | the ten figures below, as 200 dpi PNG and as vector PDF |
95
+ | `scripts/` | collector, builder, annotation tool, analysis, plotting and packaging scripts |
96
+ | `tests/` | pytest suite |
97
+
98
+ The consolidated files are produced from the raw files by
99
+ `scripts/build_dataset.py`; the raw files are never edited. No manual event
100
+ annotations were recorded during this run, so there is no `events.csv`.
101
+
102
+ ## Findings
103
+
104
+ The numbers in this section are computed by `scripts/analyze.py` and stored in
105
+ `dataset/analysis.json`.
106
+
107
+ ### Overview
108
+
109
+ ![Overview timeline](figures/png/overview_timeline.png)
110
+
111
+ 54,617 samples from 2026-09-20 12:34 to 2026-10-09 15:49 local time, 99.1 % of
112
+ the samples expected at the 30-second interval. 345 gaps are exactly 60 s,
113
+ the tick skipped after an hourly speed test; 78 gaps are longer, up to 122 s,
114
+ 53 of them after a speed test that overran two ticks and the rest after
115
+ samples whose probes waited for their timeouts. The median round-trip time
116
+ to 8.8.8.8 is 11.1 ms (95th percentile 22.3 ms), the median first-hop
117
+ round-trip time 0.2 ms, the median TCP handshake 14.5 ms.
118
+
119
+ ### Two regimes
120
+
121
+ A single least-squares change point in the daily share of samples with ICMP
122
+ loss falls on 2026-09-30 and divides the capture into two ten-day halves:
123
+
124
+ | | Sep 20 to Sep 29 | Sep 30 to Oct 9 |
125
+ |---|---|---|
126
+ | samples | 27,020 | 27,597 |
127
+ | samples with at least one lost echo | 23.2 % | 5.3 % |
128
+ | mean packet loss | 2.98 % | 0.57 % |
129
+ | round-trip time, median / p95 / p99 | 11.9 / 24.3 / 28.4 ms | 10.8 / 14.2 / 25.5 ms |
130
+ | jitter, median / p95 | 2.22 / 9.22 ms | 0.78 / 4.56 ms |
131
+ | HTTP response, median / p95 | 87.7 / 1,094 ms | 77.6 / 157 ms |
132
+ | HTTP responses slower than 1 s | 5.87 % | 1.09 % |
133
+ | recursive DNS, median | 62.5 ms | 56.2 ms |
134
+ | samples with any label | 38.9 % | 9.2 % |
135
+
136
+ ![Regimes](figures/png/regimes.png)
137
+
138
+ ![Labels](figures/png/labels.png)
139
+
140
+ ### Loss: daytime, weekdays, upstream of the gateway
141
+
142
+ ![Diurnal profile](figures/png/diurnal_profile.png)
143
+
144
+ ![Calendar heat maps](figures/png/calendar_heatmaps.png)
145
+
146
+ 14.1 % of all samples lost at least one of their ten echoes; 6,309 lost
147
+ exactly one, 1,092 two, 230 three, 41 four. On weekdays the hourly share of
148
+ samples with loss rises from 6 to 8 % between midnight and 05:00 to 25.2 % at
149
+ 08:00 and stays above 20 % until 16:00; on weekends it stays between 6 % and
150
+ 14 % all day. Weekdays as a whole have 15.2 % of samples with loss, weekends
151
+ 10.5 %. Outside the four outages, the gateway lost a packet in exactly one
152
+ sample, so the loss sits beyond the first hop.
153
+
154
+ ![Loss anatomy](figures/png/loss_anatomy.png)
155
+
156
+ Loss is only weakly bursty: a sample that follows a lossy sample is lossy with
157
+ probability 0.30 against a base rate of 0.14, and the autocorrelation of the
158
+ loss indicator is 0.18 at lag 1. The other probes show what a lost echo means:
159
+
160
+ | | no loss | 1 of 10 lost | 2 or more lost |
161
+ |---|---|---|---|
162
+ | samples | 46,907 | 6,309 | 1,383 |
163
+ | round-trip time, median | 11.0 ms | 11.8 ms | 12.6 ms |
164
+ | jitter, median | 1.00 ms | 2.12 ms | 3.29 ms |
165
+ | TCP handshake, median | 14.5 ms | 14.3 ms | 14.4 ms |
166
+ | HTTP responses slower than 1 s | 2.54 % | 7.47 % | 16.34 % |
167
+ | recursive DNS slower than 200 ms | 1.75 % | 5.22 % | 9.60 % |
168
+
169
+ The HTTP response time has a smooth body with a median of 81.9 ms and a 95th
170
+ percentile of 271 ms, then discrete modes: 2.48 % of responses fall between
171
+ 1.0 and 1.4 s and 0.55 % between 3.9 and 4.4 s. Their timing matches the
172
+ retransmission timers of the operating system's DNS client and TCP stack (one
173
+ second, then multiples of it), and they occur three to six times more often in
174
+ samples that also lost an echo, so they are lost packets waiting out a timer
175
+ rather than slow servers: the mechanism by which an echo lost in a ten-second
176
+ burst becomes a one-second stall for an application.
177
+
178
+ ### Latency episodes: path changes and a first-hop event
179
+
180
+ ![Degradation anatomy](figures/png/degradation_anatomy.png)
181
+
182
+ Runs of the high-latency label, merged across gaps of up to ten samples, give
183
+ nine episodes of 30 minutes or more; they contain 88.4 % of all high-latency
184
+ labels. Each is classified from the probe that moved: *first hop* if the
185
+ median gateway round-trip time exceeds the run's gateway threshold (5.2 ms),
186
+ *path change* if the median minimum ICMP time or the median fastest TCP
187
+ handshake rises at least 5 ms above its run median (10.0 and 12.3 ms), and
188
+ *queueing* otherwise.
189
+
190
+ | start (local) | end (local) | hours | ICMP min | ICMP mean | TCP fastest | gateway | kind |
191
+ |---|---|---|---|---|---|---|---|
192
+ | Mon Sep 21 22:34 | Tue Sep 22 01:05 | 2.52 | 20.0 | 26.4 | 25.7 | 0.0 | path change |
193
+ | Tue Sep 22 01:27 | 02:41 | 1.23 | 10.0 | 20.4 | 20.1 | 0.0 | path change |
194
+ | Tue Sep 22 02:47 | 03:44 | 0.95 | 11.0 | 21.1 | 19.5 | 0.0 | path change |
195
+ | Thu Sep 24 23:51 | Fri Sep 25 01:16 | 1.42 | 11.0 | 22.3 | 20.1 | 0.0 | path change |
196
+ | Sat Sep 26 14:07 | 14:54 | 0.78 | 14.0 | 23.5 | 18.1 | 11.4 | first hop |
197
+ | Sat Sep 26 15:01 | 15:32 | 0.52 | 14.0 | 22.0 | 18.1 | 11.2 | first hop |
198
+ | Sat Sep 26 16:36 | 18:17 | 1.68 | 14.0 | 23.7 | 20.7 | 10.6 | first hop |
199
+ | Sun Sep 27 00:00 | 06:16 | 6.28 | 17.0 | 25.6 | 23.8 | 0.0 | path change |
200
+ | Tue Sep 29 21:21 | Wed Sep 30 07:34 | 10.22 | 20.0 | 24.8 | 24.5 | 0.0 | path change |
201
+
202
+ Values are medians over the episode, in milliseconds; an end time without a
203
+ date is on the start date. The six path changes fall on four nights and all
204
+ lie between 21:21 and 07:34. In every one of them the fastest TCP handshake
205
+ sits 7 to 13 ms above its 12.3 ms baseline, the gateway is untouched, and the
206
+ DNS probe, whose queries the gateway forwards to the upstream resolver, stays
207
+ within 4 ms of its 36 ms median, so the extra delay lies on the route towards
208
+ 8.8.8.8, beyond the first hop and beyond the point where the path to the
209
+ resolver branches off. The three first-hop episodes are one event on the
210
+ afternoon of Saturday September 26: the gateway answered in 10 to 11 ms
211
+ instead of 0.2 ms, every other probe inherited the delay, and that day holds
212
+ 507 of the 560 gateway-degraded labels.
213
+
214
+ ### Outages
215
+
216
+ | start (local) | duration | samples | gateway reachable |
217
+ |---|---|---|---|
218
+ | Mon Sep 21 13:03 | 2.0 min | 4 | no |
219
+ | Thu Sep 24 07:10 | 2.5 min | 5 | no |
220
+ | Sat Sep 26 13:11 | 2.5 min | 5 | no |
221
+ | Tue Oct 6 02:51 | 2.0 min | 4 | no |
222
+
223
+ Every outage took the gateway down together with the target, with name
224
+ resolution failing as well (`getaddrinfo failed` in the `notes` column), so
225
+ all four were first-hop events. The ISP path never failed on its own during
226
+ the capture. Outages account for 18 samples, 0.03 % of the dataset.
227
+
228
+ ### The TCP/ICMP split
229
+
230
+ ![TCP versus ICMP](figures/png/tcp_vs_icmp.png)
231
+
232
+ Over the whole capture the fastest TCP handshake exceeds the minimum ICMP time
233
+ by 2.27 ms (median). From 22:00 on October 6 the share of samples in which it
234
+ exceeds it by more than 5 ms jumps from 7.2 % to 63.7 %, and the fastest
235
+ handshake becomes bimodal (first quartile 12.9 ms, median 18.6 ms, third
236
+ quartile 22.2 ms) while the ICMP minimum stays at 10 ms. The ICMP and TCP
237
+ probes address the same host, so the difference is in how the two kinds of
238
+ flow are routed; a monitor that relies on ping alone would not see the change.
239
+
240
+ ### DNS
241
+
242
+ The "cached" probe asks the gateway resolver for `example.com`, which it has
243
+ answered before, yet its median is 36.0 ms (5th percentile 17 ms; only 3.4 %
244
+ of answers arrive within 15 ms). The gateway therefore forwards every query
245
+ upstream rather than serving it from a local cache, and the column measures
246
+ the upstream resolver's responsiveness. The recursive probe (a random name
247
+ under `example.com`) has a median of 59.2 ms, 20.2 ms above the cached probe
248
+ in the same sample; 2.35 % of recursive lookups take longer than 200 ms and
249
+ the 99th percentile is 1,266 ms. 24 samples recorded a DNS timeout, 14 of
250
+ them during the four outages (the other four outage samples failed with a
251
+ socket error instead).
252
+
253
+ ### Memory
254
+
255
+ ![Autocorrelation](figures/png/autocorrelation.png)
256
+
257
+ | lag | 30 s | 5 min | 30 min | 3 h | 12 h | 24 h |
258
+ |---|---|---|---|---|---|---|
259
+ | log round-trip time | 0.81 | 0.76 | 0.68 | 0.42 | 0.18 | 0.09 |
260
+ | jitter | 0.44 | 0.42 | 0.36 | 0.25 | 0.23 | 0.17 |
261
+ | lost-echo indicator | 0.18 | 0.17 | 0.16 | 0.14 | 0.04 | 0.10 |
262
+ | fastest TCP handshake | 0.28 | 0.15 | 0.14 | 0.09 | 0.04 | 0.02 |
263
+ | recursive DNS | 0.04 | 0.02 | 0.02 | 0.02 | 0.00 | 0.01 |
264
+ | HTTP | 0.06 | 0.05 | 0.03 | 0.04 | 0.02 | 0.01 |
265
+
266
+ Latency is persistent for hours and has no daily cycle at this site; DNS and
267
+ HTTP timings are memoryless at every lag.
268
+
269
+ ### WAN throughput
270
+
271
+ ![Speed tests](figures/png/speedtest.png)
272
+
273
+ 460 hourly speed tests ran through the `speedtest-cli` Python fallback with no
274
+ pinned server and used 162 different servers. 79 tests could not measure the
275
+ server's latency (the fallback encodes that failure as a value of 600,000 ms
276
+ or more; the builder sets it to missing), 7 reported a download of 0 Mbit/s
277
+ and 4 an upload of 0 Mbit/s. Download: median 34.7 Mbit/s, 90th percentile
278
+ 183 Mbit/s, maximum 480.7 Mbit/s; among the ten most used servers the
279
+ per-server median ranges from 20.9 to 371.8 Mbit/s. Upload: median
280
+ 28.6 Mbit/s, 90th percentile 40.3 Mbit/s, maximum 54.2 Mbit/s. The median
281
+ download is 61.5 Mbit/s for the 95 tests started between 00:00 and 04:59 and
282
+ 22.0 Mbit/s for the 156 started between 09:00 and 16:59. The download column is a lower bound on
283
+ the access link and depends heavily on the server; the upload column is more
284
+ consistent. Loaded latency was not measured because the official Ookla CLI was
285
+ not installed.
286
+
287
+ ## Measurement protocol
288
+
289
+ One sample is taken at every tick of a fixed schedule (`start + k * interval`).
290
+ Within a sample the following probes run concurrently:
291
+
292
+ - ICMP: `ping_count` (10) echo requests to the target and `gateway_ping_count`
293
+ (5) to the default gateway, one second apart, each waiting at most
294
+ `ping_timeout` (2 s) for its reply. Per-reply round-trip times are parsed
295
+ from the output of the operating system's `ping`.
296
+ - TCP: `tcp_count` (5) TCP handshakes to `target:443`, 0.2 s apart, timed with
297
+ a high-resolution clock from `connect()` start to completion.
298
+ - DNS: two UDP queries (type A, recursion desired) sent directly to the
299
+ configured resolver, bypassing the operating-system cache. The first asks for
300
+ `dns_host` (`example.com`), which the resolver serves from its cache after
301
+ the first sample and therefore measures the resolver's responsiveness. The
302
+ second asks for a random 12-hexadecimal-character subdomain of `dns_host`,
303
+ which can never be cached and therefore measures a full recursive
304
+ resolution. A response with rcode NOERROR or NXDOMAIN is a valid measurement.
305
+ - HTTP: one GET request for `http_url`
306
+ (`http://www.msftconnecttest.com/connecttest.txt`) with the system proxy
307
+ bypassed, timed until the response headers have been received.
308
+ - Wi-Fi: link metrics read from the operating system
309
+ (`netsh wlan show interfaces`, `iw dev <if> link`, `system_profiler
310
+ SPAirPortDataType`).
311
+
312
+ After the concurrent probes, and only on their own cadence (every hour in the
313
+ published run), the heavy tests run: an Ookla speed test and/or an `iperf3`
314
+ test. Because they run after the latency probes of the same sample, they do
315
+ not influence that sample's latency columns; a sample that includes a speed
316
+ test usually overruns the interval, and the collector then skips to the next
317
+ tick.
318
+
319
+ The default gateway, the interface type and the resolver are detected at start
320
+ and re-detected every 15 minutes (or within a minute after the gateway probe
321
+ loses every packet). System sleep is inhibited for the duration of the run.
322
+
323
+ Run configuration of the published capture: interval 30 s, 10 echoes to
324
+ `8.8.8.8` and 5 to the gateway with a 2 s timeout, 5 TCP handshakes to port
325
+ 443, DNS probes against the gateway resolver with a 3 s timeout, HTTP probe
326
+ with a 5 s timeout, Wi-Fi metrics enabled, SSID not recorded, speed test every
327
+ hour with no pinned server, no iperf3 server, collector version 2.0.0.
328
+
329
+ ## Schema
330
+
331
+ Empty cells mean "not measured or failed"; the `notes` column records the
332
+ reason (for example `ping: no replies`, `dns recursive: timeout`,
333
+ `speedtest (ookla) failed: ...`). Times are in milliseconds unless the column
334
+ name says otherwise.
335
+
336
+ ### Identity and time
337
+
338
+ | Column | Type | Description |
339
+ |---|---|---|
340
+ | `run_id` | string | collection run identifier, `<host_label_>YYYYMMDD_HHMMSS` in local time of the run start |
341
+ | `host_label` | string | free-text label of the vantage point (`--host-label`), empty if not given |
342
+ | `sample_index` | int | 0-based row number within the run |
343
+ | `timestamp_utc` | string | sample start, ISO 8601 UTC, e.g. `2026-09-05T14:00:00Z` |
344
+ | `timestamp_local` | string | sample start in the collector's local time with UTC offset, e.g. `2026-09-05T10:00:00-04:00` |
345
+ | `epoch_s` | int | sample start as Unix time (seconds) |
346
+ | `elapsed_s` | float | seconds since the start of the run |
347
+ | `link_type` | string | `wifi`, `ethernet`, `other` or `unknown`: the interface that carries the default route |
348
+ | `gateway_ip` | string | default gateway (private address) |
349
+ | `dns_server` | string | resolver used for the DNS probes |
350
+ | `target` | string | host used for the ICMP and TCP probes |
351
+
352
+ ### Run context (added by the builder)
353
+
354
+ | Column | Type | Description |
355
+ |---|---|---|
356
+ | `collector_version` | string | version of `collect_data.py` that wrote the run |
357
+ | `collector_os` | string | `Windows`, `Linux` or `macOS` |
358
+ | `interval_s` | float | configured sampling interval of the run |
359
+ | `rtt_resolution_ms` | float | resolution of the ICMP columns: 1.0 on Windows, 0.001 elsewhere |
360
+
361
+ ### Time features (added by the builder, from `timestamp_local`)
362
+
363
+ | Column | Type | Description |
364
+ |---|---|---|
365
+ | `date_local` | string | local calendar date `YYYY-MM-DD` |
366
+ | `hour_local` | int | local hour 0-23 |
367
+ | `weekday_local` | int | 0 = Monday ... 6 = Sunday |
368
+ | `is_weekend` | int | 1 on Saturday and Sunday |
369
+ | `gap_before_s` | float | seconds since the previous sample of the same run; larger than `interval_s` after a skipped tick or a sleep/outage gap |
370
+
371
+ ### Target latency (ICMP)
372
+
373
+ | Column | Type | Description |
374
+ |---|---|---|
375
+ | `rtt_min_ms` | float | minimum round-trip time of the received replies |
376
+ | `rtt_avg_ms` | float | arithmetic mean |
377
+ | `rtt_median_ms` | float | median |
378
+ | `rtt_max_ms` | float | maximum |
379
+ | `rtt_stddev_ms` | float | population standard deviation of the received replies (the `mdev` of Linux `ping`) |
380
+ | `jitter_ms` | float | mean absolute difference between consecutive received replies (packet delay variation in the sense of RFC 3393, applied to round-trip times); empty when fewer than two replies were received |
381
+ | `packets_sent` | int | echo requests sent (`ping_count`) |
382
+ | `packets_received` | int | replies received |
383
+ | `packet_loss_pct` | float | `100 * (sent - received) / sent` |
384
+
385
+ ### Gateway latency (ICMP to the default gateway)
386
+
387
+ | Column | Type | Description |
388
+ |---|---|---|
389
+ | `gateway_rtt_min_ms` | float | minimum round-trip time |
390
+ | `gateway_rtt_avg_ms` | float | mean |
391
+ | `gateway_rtt_max_ms` | float | maximum |
392
+ | `gateway_jitter_ms` | float | mean absolute consecutive difference |
393
+ | `gateway_packet_loss_pct` | float | loss in percent |
394
+
395
+ ### TCP, DNS and HTTP
396
+
397
+ | Column | Type | Description |
398
+ |---|---|---|
399
+ | `tcp_connect_min_ms` | float | fastest of the TCP handshakes to `target:443` |
400
+ | `tcp_connect_median_ms` | float | median handshake time |
401
+ | `dns_resolve_ms` | float | response time of the resolver for `dns_host` (cache-served after the first sample) |
402
+ | `dns_recursive_ms` | float | response time for a random subdomain of `dns_host` (full recursive lookup) |
403
+ | `http_ms` | float | time to the response headers of `http_url`, including the operating-system name lookup, the TCP handshake and the request |
404
+ | `http_status` | int | HTTP status code of that response |
405
+
406
+ ### Wi-Fi link (empty on wired links)
407
+
408
+ | Column | Type | Description |
409
+ |---|---|---|
410
+ | `wifi_ssid` | string | network name; only recorded with `--record-ssid` |
411
+ | `wifi_signal_pct` | int | link quality 0-100: the value reported by Windows; on Linux and macOS derived from RSSI with the Windows mapping `quality = 2 * (rssi + 100)` clipped to 0-100 |
412
+ | `wifi_rssi_dbm` | float | received signal strength: as reported on Linux and macOS; on Windows derived from the quality with `rssi = quality / 2 - 100` |
413
+ | `wifi_rx_mbps` | float | negotiated receive PHY rate (Mbit/s) |
414
+ | `wifi_tx_mbps` | float | negotiated transmit PHY rate (Mbit/s); equals the receive rate on macOS, which reports one rate |
415
+ | `wifi_channel` | int | channel number |
416
+ | `wifi_band_ghz` | float | 2.4, 5 or 6 |
417
+ | `wifi_phy` | string | standard in use, e.g. `802.11ax` |
418
+
419
+ ### WAN throughput (rows in which a speed test ran)
420
+
421
+ | Column | Type | Description |
422
+ |---|---|---|
423
+ | `download_mbps` | float | download throughput (Mbit/s) |
424
+ | `upload_mbps` | float | upload throughput (Mbit/s) |
425
+ | `speedtest_ping_ms` | float | idle latency to the speed-test server as reported by the tool; missing when the `speedtest-cli` fallback could not measure it |
426
+ | `download_loaded_latency_ms` | float | interquartile-mean latency measured during the download (Ookla CLI only) |
427
+ | `upload_loaded_latency_ms` | float | interquartile-mean latency measured during the upload (Ookla CLI only) |
428
+ | `speedtest_server_id` | int | Ookla server id |
429
+ | `speedtest_server` | string | server name and location |
430
+
431
+ ### LAN throughput (rows in which an iperf3 test ran)
432
+
433
+ | Column | Type | Description |
434
+ |---|---|---|
435
+ | `lan_throughput_mbps` | float | TCP throughput to the LAN iperf3 server (Mbit/s) |
436
+ | `lan_retransmits` | int | TCP retransmissions during the test, when reported |
437
+
438
+ ### Reference labels (added by the builder)
439
+
440
+ | Column | Type | Rule |
441
+ |---|---|---|
442
+ | `label_outage` | int | `packets_received == 0` while `packets_sent > 0` |
443
+ | `label_packet_loss` | int | `0 < packet_loss_pct < 100` |
444
+ | `label_high_latency` | int | `rtt_avg_ms > T_rtt` |
445
+ | `label_high_jitter` | int | `jitter_ms > T_jitter` |
446
+ | `label_gateway_degraded` | int | `gateway_rtt_avg_ms > T_gateway` or `gateway_packet_loss_pct > 0` |
447
+ | `label_anomaly` | int | any of the above |
448
+
449
+ The thresholds are computed per run from that run's own distribution with a
450
+ robust rule, so they adapt to the baseline of the site without being pulled by
451
+ the outliers they are meant to find:
452
+
453
+ ```text
454
+ T = median + max(k * 1.4826 * MAD, absolute floor, relative floor * median)
455
+ ```
456
+
457
+ with `k = 5`, the median absolute deviation `MAD`, and the floors
458
+ 10 ms / 50 % for `rtt_avg_ms`, 5 ms / 100 % for `jitter_ms` and
459
+ 5 ms / 100 % for `gateway_rtt_avg_ms`. A threshold is not computed (and the
460
+ label stays 0) for runs with fewer than 20 valid values. For the published run
461
+ the thresholds are 21.1 ms (`rtt_avg_ms`), 6.0 ms (`jitter_ms`) and 5.2 ms
462
+ (`gateway_rtt_avg_ms`); the values used for each run are stored in
463
+ `dataset/summary.json`. These labels are derived from the measurements
464
+ themselves; they are reference points for evaluating detectors, not
465
+ independent ground truth.
466
+
467
+ ### Manual annotations and notes
468
+
469
+ | Column | Type | Description |
470
+ |---|---|---|
471
+ | `event_category` | string | categories of the manual events covering the sample time, joined with ` \| `; one of `load`, `outage`, `maintenance`, `environment`, `isp`, `device`, `other` |
472
+ | `event_description` | string | the corresponding free-text descriptions |
473
+ | `sample_duration_s` | float | wall-clock time the sample's probes took |
474
+ | `notes` | string | probe errors and status messages, `; `-separated |
475
+
476
+ ## Validation performed by the builder
477
+
478
+ - Rows with a malformed `timestamp_utc` are dropped and duplicate
479
+ `(run_id, timestamp_utc)` pairs are reduced to the first occurrence (none in
480
+ the published run).
481
+ - Loss percentages outside 0-100 are set to missing (none in the published
482
+ run).
483
+ - `speedtest_ping_ms` values of 600,000 ms or more are set to missing. The
484
+ `speedtest-cli` fallback scores each failed latency attempt as 3,600 s and
485
+ averages three attempts over six, so one failed attempt already yields
486
+ 600,000 ms and three yield 1,800,000 ms. The raw file keeps the original
487
+ values; 79 of the 460 speed-test rows are affected. Collector version 2.0.1
488
+ no longer records such values.
489
+
490
+ ## Statistics
491
+
492
+ <!-- BEGIN GENERATED STATISTICS -->
493
+ Generated by `scripts/build_dataset.py` on 2026-10-10T05:16:24Z from 1 collection run(s).
494
+
495
+ | Item | Value |
496
+ |---|---|
497
+ | Rows | 54,617 |
498
+ | Columns | 67 |
499
+ | First sample (UTC) | 2026-09-20T16:34:22Z |
500
+ | Last sample (UTC) | 2026-10-09T19:49:43Z |
501
+ | Span | 19.14 days |
502
+ | Sampling coverage | 99.1 % of the samples expected at the configured interval |
503
+ | Manual event annotations | 0 |
504
+
505
+ | Run | Vantage | Link | OS | Interval | Rows | Span | Coverage |
506
+ |---|---|---|---|---|---|---|---|
507
+ | desktop-eth_20260920_123407 | desktop-eth | unknown (54617) | Windows | 30 s | 54,617 | 459.26 h | 99.1 % |
508
+
509
+ | Metric | n | Median | Mean | p95 | Max |
510
+ |---|---|---|---|---|---|
511
+ | rtt_avg_ms | 54,599 | 11.1 | 12.581 | 22.3 | 288.6 |
512
+ | rtt_min_ms | 54,599 | 10.0 | 10.432 | 14.0 | 263.0 |
513
+ | rtt_max_ms | 54,599 | 13.0 | 17.869 | 32.0 | 373.0 |
514
+ | jitter_ms | 54,599 | 1.0 | 2.306 | 7.667 | 59.778 |
515
+ | packet_loss_pct | 54,617 | 0.0 | 1.763 | 10.0 | 100.0 |
516
+ | gateway_rtt_avg_ms | 54,599 | 0.2 | 0.321 | 0.8 | 89.4 |
517
+ | gateway_packet_loss_pct | 54,617 | 0.0 | 0.033 | 0.0 | 100.0 |
518
+ | tcp_connect_median_ms | 54,598 | 14.484 | 19.311 | 32.127 | 1524.463 |
519
+ | dns_resolve_ms | 54,595 | 36.026 | 43.594 | 74.878 | 3252.214 |
520
+ | dns_recursive_ms | 54,590 | 59.242 | 87.275 | 106.31 | 25645.041 |
521
+ | http_ms | 54,565 | 81.911 | 154.745 | 271.36 | 13111.02 |
522
+ | wifi_signal_pct | 4 | 85.0 | 85.75 | 88.4 | 89.0 |
523
+ | wifi_rssi_dbm | 4 | -57.5 | -57.125 | -55.8 | -55.5 |
524
+ | wifi_rx_mbps | 4 | 866.7 | 866.7 | 866.7 | 866.7 |
525
+ | download_mbps | 460 | 34.725 | 68.463 | 318.275 | 480.72 |
526
+ | upload_mbps | 460 | 28.625 | 26.121 | 41.142 | 54.17 |
527
+ | speedtest_ping_ms | 381 | 52.058 | 93.665 | 196.381 | 1167.189 |
528
+
529
+ | Reference label | Rows | Share |
530
+ |---|---|---|
531
+ | label_outage | 18 | 0.033 % |
532
+ | label_packet_loss | 7,692 | 14.084 % |
533
+ | label_high_latency | 3,067 | 5.615 % |
534
+ | label_high_jitter | 5,074 | 9.29 % |
535
+ | label_gateway_degraded | 560 | 1.025 % |
536
+ | label_anomaly | 13,036 | 23.868 % |
537
+
538
+ <!-- END GENERATED STATISTICS -->
539
+
540
+ ## Known limitations and biases
541
+
542
+ - Single vantage point, single run. Absolute values reflect one site's access
543
+ technology, router, cabling and ISP routing. Relative variation, the
544
+ structure of episodes and the probe-to-probe relationships transfer better
545
+ than absolute baselines.
546
+ - `link_type` is `unknown` for every row. The PowerShell route query returned
547
+ nothing on this machine, the collector fell back to `route print`, and that
548
+ fallback does not report the media type. The vantage point is a wired
549
+ desktop (host label `desktop-eth`, median first-hop round-trip time 0.2 ms).
550
+ Four rows carry Wi-Fi link metrics (channel 44 or 157, 5 GHz, 802.11ac,
551
+ 866.7 Mbit/s): a wireless adapter on the same machine was connected at those
552
+ moments. Their first-hop round-trip times (0.0 to 1.2 ms) and target
553
+ latencies match the wired baseline, so the default route stayed on the
554
+ wired link.
555
+ - ICMP resolution on Windows is 1 ms and sub-millisecond replies are stored as
556
+ 0; jitter and standard deviation computed from such values contain
557
+ quantisation noise, and the gateway columns are mostly 0 for this wired
558
+ site. Use `tcp_connect_*` for sub-millisecond latency, or collect on
559
+ Linux/macOS. `rtt_resolution_ms` states the resolution of each run.
560
+ - ICMP is handled differently from user traffic by some routers and hosts
561
+ (rate limiting, de-prioritisation). The TCP connect and HTTP columns measure
562
+ the transport and application paths that applications actually use, and the
563
+ October 6 TCP/ICMP split shows that the two can be routed differently.
564
+ - The 10-echo burst characterises latency over roughly ten seconds of each
565
+ interval; events shorter than that between samples are not observed.
566
+ - The DNS "cached" probe is cached at the upstream resolver, not at the
567
+ gateway: at this site the gateway forwards every query, so the column
568
+ measures the upstream resolver plus the forwarding hop.
569
+ - The DNS "recursive" probe generates NXDOMAIN answers for random names; a
570
+ resolver that applies aggressive negative caching or that intercepts
571
+ NXDOMAIN (ISP search redirection) still answers, but its timing reflects that
572
+ behaviour.
573
+ - The HTTP probe includes the operating system's name lookup for the URL's
574
+ host, which is normally cached; its discrete modes near 1.1, 1.3 and 4.1 s
575
+ are retransmission timers waiting out lost packets, not propagation or
576
+ queueing delay.
577
+ - Speed tests were run with the `speedtest-cli` fallback, which uses the
578
+ Python interpreter's own HTTP transfers, and without a pinned server. The
579
+ download column is a lower bound that depends strongly on the chosen
580
+ server; 79 tests have no latency, 7 report 0 Mbit/s download and 4 report
581
+ 0 Mbit/s upload. Loaded latency is absent. Speed tests also load the
582
+ household link for about half a minute, which is visible to other devices,
583
+ and they leave a 60-second gap in the 30-second grid.
584
+ - The two halves of the capture differ markedly (see the findings). Models
585
+ trained on one half should be evaluated on the other.
586
+ - No manual event annotations were recorded, so the `event_*` columns are
587
+ empty and the causes of the episodes are inferred from the probes only.
588
+ - Gaps: system sleep, reboots and collector restarts appear as gaps
589
+ (`gap_before_s`) or as separate runs; ticks skipped after an overrun also
590
+ appear as gaps. The published run has none longer than 122 s.
591
+ - Clock: timestamps come from the system clock; `elapsed_s` uses a monotonic
592
+ clock and is unaffected by clock adjustments.
593
+
594
+ ## Privacy
595
+
596
+ The dataset does not contain public IP addresses, hostnames of household
597
+ devices, MAC addresses, Wi-Fi network names or per-device traffic. Gateway and
598
+ resolver addresses are private LAN addresses. The run metadata records the
599
+ operating system build, the Python version, the time zone and the private LAN
600
+ address of the collector. Speed-test rows name the test server, which was not
601
+ pinned and therefore only indicates the broad region of the site.
602
+
603
+ ## Citation
604
+
605
+ If you use the data:
606
+
607
+ ```bibtex
608
+ @dataset{ezharjan_localnetdynamics_data_2026,
609
+ author = {Ezharjan, Alexander},
610
+ title = {LocalNet-Dynamics: A Longitudinal Local Area Network Telemetry Dataset},
611
+ year = {2026},
612
+ version = {1.0.0},
613
+ publisher = {Hugging Face},
614
+ doi = {10.57967/hf/10841},
615
+ url = {https://huggingface.co/datasets/ezharjan/LocalNetDynamics}
616
+ }
617
+ ```
618
+
619
+ If you use the collector, the build pipeline or the analysis code:
620
+
621
+ ```bibtex
622
+ @software{ezharjan_localnetdynamics_code_2026,
623
+ author = {Ezharjan, Alexander},
624
+ title = {LocalNet-Dynamics: collector, build pipeline and analysis toolkit for longitudinal home-network telemetry},
625
+ year = {2026},
626
+ version = {1.0.0},
627
+ url = {https://github.com/Ezharjan/LocalNetDynamicsDataset}
628
+ }
629
+ ```
scripts/package_huggingface.py CHANGED
@@ -40,6 +40,7 @@ authors:
40
  version: {version}
41
  date-released: "{date}"
42
  license: MIT
 
43
  url: "https://huggingface.co/datasets/ezharjan/LocalNetDynamics"
44
  repository-code: "https://github.com/Ezharjan/LocalNetDynamicsDataset"
45
  abstract: >-
 
40
  version: {version}
41
  date-released: "{date}"
42
  license: MIT
43
+ doi: 10.57967/hf/10841
44
  url: "https://huggingface.co/datasets/ezharjan/LocalNetDynamics"
45
  repository-code: "https://github.com/Ezharjan/LocalNetDynamicsDataset"
46
  abstract: >-