Souk's Notes - Unofficial Summaries

2022-06-24 · open the original ↗

• Content shared in web meetings is largely inaccessible to blind attendees. Jones, of Microsoft's commercial software engineering team, says shared screens arrive as a video stream that strips the document's structure, so screen readers cannot follow it in real time. VocalScreen was the idea of a former colleague, Hassan, an engineer blind from birth; volunteer crews have iterated on it in one-week projects for more than two years.
• Presenters will only use assistive tools that start in seconds. The first version was a Windows desktop app built on Windows APIs, but Jones says nobody can hold up a meeting five or ten minutes to install software. The latest version is an Edge and Chrome extension for the presenter, who posts a unique session link in the chat; listeners need only a web browser, with an optional extension for extra features.
• Streaming the whole HTML DOM preserves what screen readers need. Jones says sending only the text would lose navigation and the relationships between elements, so it sends the DOM with ARIA labels and H1/H2 headers intact. He cites a recent Starbucks accessibility project where ARIA labels announced events a sighted user sees, such as a settings menu opening.
• Less structured content would get machine-generated structure. On the roadmap, Jones lists an Office extension (PowerPoint being perhaps the most shared content), eventually Teams, and cloud OCR that would use text size and position to generate ARIA-style labels with machine learning. It is an internal "hack for good" project, not public as of recording; the hope is to feed the features to the Teams product group, and it may be open-sourced someday.
• People with disabilities are some of the best problem solvers around. Jones traces his involvement to a lunch with Microsoft's chief accessibility officer that he bid on in the company's annual Give campaign. He urges technologists to get involved, calling accessibility both a human cause and a place to work with cutting-edge technology.

2022-06-21 · open the original ↗

• Install Node from nodejs.org, which detects your operating system. The site offers two downloads: the latest stable release, and the latest release, which may have bugs. Giard says he usually takes the stable one. A downloads tab lists every release.
• A web server takes a handful of lines using Node's http library. In a new file in an empty folder in Visual Studio Code (the file can have any name ending in .js), he imports the http library into a variable named http. He then calls its create-server method with a function taking request and response arguments (req, res) and has it listen on port 8081. Whatever that function returns is what a user hitting the port receives. Inside it, res.writeHead sets a 200 OK status and a text/plain content type, res.write outputs "hello world", and res.end tells the browser that no more content is coming. He also logs a "server running" message, as a reminder of where to browse.
• Run it from VS Code's debugger or the command line. Run and Debug printed "server running at localhost:8081", and the page loaded in a browser. Once he stopped it, the page just spun and timed out. VS Code is not required. The same thing runs as node.exe followed by the path to app.js, from PowerShell for example, or from a script. The full path to node.exe is only needed if Node is not on your PATH.
• Leave a previous instance running and the port is already in use. His second run failed because the PowerShell copy was still running. Stopping it with Ctrl+C cleared the error.
• The response can carry HTML and output generated by JavaScript. He added heading and paragraph tags to the output, then a for loop that writes the numbers 1 through 10 as list items inside a ul. The pasted loop at first failed because it used response rather than the argument's name, res. His advice: think about what you paste.

2022-06-20 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2022-06-19 · open the original ↗

• Data engineers still protect data lakes by hand-copying data. Polak, who runs developer experience at the company behind LakeFS, says pipelines on object stores such as S3 or Azure Blob are transformed while consumers are already reading them, so engineers copy and snapshot petabytes manually. Her example: copy 32 tables, and statistically about two of them get messed up, leaving the business to explain corrupted data.
• LakeFS applies Git's branch, merge and rollback to data. Engineers branch off production data to experiment or debug against its real volume and variety, then merge back or roll back, and can travel to a past timestamp across the repository. She stresses, against the host's framing that it "sits on top of" the store, that it tracks only metadata about what changed rather than piping the data itself, and works on binary files, so any format (Parquet, Avro, CSV, images, video) is supported.
• Pre-merge hooks let teams enforce data quality programmatically. Modelled on GitHub Actions, a hook can block a data branch from merging until checks pass. Her example: a data scientist sets an expected gender distribution, so an ETL that silently dropped "corrupted" records, which turned out to be all the female records, gets caught before biased data reaches analytics or machine learning.
• The open-source version is complete; the paid product is management. Polak says the core engine, UI, CLI and SDK are free under Apache 2.0 with no functionality withheld, and the commercial LakeFS Cloud, then launching, sells a managed, scalable deployment to companies that don't want to run a distributed system themselves. She claims more than 1,000 companies use it and more than 4,000 community members; docs and community are at lakefs.io.
• Data lakes lack lifecycle management that databases have long had. She says her team is working on content and an architecture for managing production data over object stores in a scalable, secure way. She also points to a 2022 data-engineering landscape article written with other companies.

2022-06-17 · open the original ↗

• ADX ran Microsoft's own telemetry for years before customers got it. Rauch, a product manager on the team since 2019, says it began about 10 years ago as a four-person startup in Israel built to handle Power BI log and telemetry data ("Kusto"), and opened to external customers in February 2019. He claims about 99.9% of Microsoft's log and telemetry data lives in it, that it ingests more than 100 petabytes a day, and that clusters hold about 8 exabytes in total; Azure Monitor, Sentinel and IoT Central are built on it.
• Dropping transactions buys ingest speed and scale. ADX is an append-only analytical database with no ACID guarantees, which Rauch says allows higher ingest bandwidth. Lauzon cites the tagline "billion queries over a billion records under 1 second," says customers typically arrive after another database broke or became too expensive at scale, and describes clusters of up to 1,000 nodes with predictive autoscaling that uses machine learning to scale before peaks. Both say it is wrong for media storage, transactional backends, model training (it only scores imported models) and long ETL, and queries have a hard one-hour cap.
• Time series handling is the feature that turned customers. Rauch describes per-table retention periods, a "hot" window cached on local SSDs over Azure storage, and automatic partitioning by ingestion time. The make-series operator returns vectors rather than rows, enabling interpolation of missing IoT readings, forecasting, anomaly detection and regression without deploying a machine learning model; he calls this the point where users saw it as more than cheap log storage.
• Getting started needs only a Microsoft identity. Lauzon says to load data and start querying: a free cluster requires no Azure subscription or credit card, and one-click ingestion infers schema and mapping from Parquet, CSV or Avro files. Connectors cover Event Hub, IoT Hub, Kafka, Spark and Logstash, and each cluster exposes a T-SQL endpoint (port 1433, Rauch thinks) that translates SQL into KQL, which Giard says helped him learn KQL.
• The roadmap extends ADX toward graphs, geospatial work and local development. As of recording, Lauzon says a Docker-based Kusto emulator for local development and CI testing arrives "next week" and Cosmos DB sync is in limited private preview. Rauch lists a new L v3 VM SKU made for Kusto, graph-query syntax in KQL, geospatial clustering (geohash, S2, H3) on top of about 40 geo functions, dashboards going GA, and customer-managed keys per database rather than per cluster, with up to 10,000 databases per cluster.

2022-06-15 · open the original ↗

• No transcript text exists for this clip. The transcript file is empty, so the 14-second short titled only "June 15, 2022" offers nothing to summarize.

2022-06-15 · open the original ↗

• The command palette lists every Terminal action and its hotkey. Hanselman's pitch, addressed to people frustrated with configuring Windows Terminal, is that Ctrl+Shift+P shows each available command and what it is bound to, which surfaces actions users never knew existed; his example is find, bound to Ctrl+Shift+F, which searches and highlights text in the buffer.
• Panes and tabs can work like a browser. Ctrl+T opens a tab and Ctrl+W closes it, alt-clicking a profile in the dropdown splits the current window into a new pane, and shift-click opens a new window. He points out that Ctrl+W for close pane and Ctrl+Shift+minus for split pane are his own bindings, not the defaults.
• Actions live both in the settings UI and the JSON file. The Actions page in settings lets you pick a command and type a preferred hotkey; the same bindings are serialized under actions in settings.json, including commands with parameters such as split pane automatic versus horizontal or vertical. He calls the JSON the better route for advanced users because the file can be kept in a GitHub gist or a cloud drive and reused on every machine.
• Editing in VS Code gets schema completion and a safety net. The Terminal's JSON schema gives VS Code a completion list of every command, and he says Terminal will not accept a saved file with bad syntax but reverts to a backup. He demonstrates by adding close other tabs on Ctrl+Alt+Shift+W in the JSON, which then shows up in the UI.

2022-06-14 · open the original ↗

• Starlink for RVs lets waitlisted buyers skip the residential queue. In a waitlist area in Idaho, the host had the same rectangular dish ordered, delivered and running in under a week; the trade-off, he says, is that RV traffic gets the lowest priority on the network, so nearby residential customers come first and speeds depend on local saturation and time of day.
• The RV plan costs the same as residential plus portability. The dish is $599 plus tax and $50 shipping, with the first month of $135 paid up front; the host equates that to the $110-a-month residential service plus its $25 portability charge.
• Its one advantage over residential is that it can be paused. The host suggests it for a seasonal vacation home, where service can be turned off for the months you are away instead of paying the $110 residential base rate year-round.
• Under trees and at lowest priority, it still delivered usable speeds. After a plug-in setup, the dish found signal and self-oriented in about five to ten minutes; three tests from 15 feet from the included router gave 64/10, 197/14 and 122/14 Mbps, and a day of remote work showed no slowdowns. His verdict: with no other internet, a minimum of about 60 Mbps is well worth the cost.

2022-06-13 · open the original ↗

• XRP holders should know Ripple's origins before buying. The speaker says Ripple's technology dates to 2004, five years before Bitcoin, and is not trustless because payments ripple through layers of trust. The current Ripple emerged in 2012 under Chris Larsen and Jed McCaleb, a Mt. Gox founder, and was then called OpenCoin; the speaker owns some.
• Ripple drifted from its creator's person-to-person vision. The original creator (Fugger, as the captions render it) wanted person-to-person payments, but XRP became bank-to-bank. The speaker says XLM fits that original vision more closely.

2022-06-13 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2022-06-11 · open the original ↗

• A prolonged sense of danger drove pandemic burnout. Sexton says she withdrew even from virtual contact, and points to psychologists' later articles arguing that fight-or-flight sustained for two years causes burnout, brain fog and an inability to answer messages. Giard adds his own view that virtual contact maintains relationships but real ones happen in person.
• A game gave her a break from what she couldn't control. Calling herself an empath who felt the weight of the world's suffering, Sexton says advice to "focus on what you can control" made her feel powerless when she wanted to fight politically and socially. She used Animal Crossing: New Horizons, released on the Nintendo Switch at the start of the pandemic, to recentre herself.
• The game's real-time clock rebuilt her relationship with time. Because the game follows the device's clock and some events are locked by Nintendo, missed days are really missed; she now knows her daily routine there takes about half an hour, which she says helped her time management. Its villagers notice when you stop playing, and two years of keeping up with them, she says, helped her get over a fear of commitment.
• Shared avatars make conversation easier than a bare phone call. Friends fly to her island over the internet while they talk on speakerphone, walk around, or sit in the in-game coffee shop; she finds audio-only calls harder to parse and more awkward. She believes up to about eight people can gather on one island.
• A mild case after vaccination made her less afraid of COVID. Sexton says catching it a couple of weeks before recording felt like a cold thanks to vaccines and boosters, though she still masks outdoors when she can. Giard says he also had a mild case after four shots and manages risk rather than eliminating it; he also describes remote therapy with a therapist in Nevada, and Sexton says therapy should be "really just a requirement."

2022-06-10 · open the original ↗

• Start with a simple model while labelled data accumulates. The host recommends beginning with something like logistic regression on a handful of hand-curated features, which also confirms whether the data collected so far carries any valuable signal. This is the third part of his three-part series on getting business value from machine learning.
• Deep learning pays off only once data is plentiful. He says the threshold depends on the problem and its signal-to-noise ratio, but generally deep learning needs at least tens of thousands of data points to be meaningfully used, and some problems need millions or, in extreme cases, billions.
• In production, speed usually matters more than accuracy. Users expect results in seconds or less, so if a large deep learning model adds only a few percent of ROC AUC over a regression but takes 10 seconds instead of one, the host argues the regression may be the better choice. In his view the accuracy gain may be imperceptible to users, but the wait will not be.

2022-06-07 · open the original ↗

• The U6 In-Wall matches full-size UniFi APs on radio power. Earlier in-wall units were weaker hotel/dorm devices, but the host puts this Wi-Fi 6 model on par with the U6 Pro: 4x4 MU-MIMO with OFDMA on 5 GHz, 2x2 MU-MIMO on 2.4 GHz, 22 dBm (2.4 GHz) and 26 dBm (5 GHz) max power, and 300+ concurrent clients per the spec he cites.
• PoE passthrough needs PoE+ in, not standard PoE. It has a rear PoE-in port plus four bottom ports, three gigabit and one 802.3af PoE out for a camera or VoIP phone; 802.3af in powers the AP but gives no PoE out, which requires 802.3at PoE+ in. It is indoor-only, rated −30 to 60 °C.
• Park replaced gear in UniFi's default site instead of unadopting it. When swapping out his UAP In-Wall HD (on UniFi 7.1.66), the host moved the old unit to the default site as a placeholder, avoiding a factory reset and re-adoption later; this works only on multi-site controllers like his hosted UniFi, not Dream Machines.
• The old UAP In-Wall HD backplate cannot be reused. The U6 In-Wall is slightly wider and uses a metal backplate instead of plastic, so the plate must be replaced; the host likes the simpler two-piece design versus the HD's three pieces plus optional cover.
• It drops the security screw and the decorative skins. The HD had a Torx security screw; the U6 relies on a click lock the host says a paper clip can defeat, a loss for hotels and dorms. No skins exist even in the early access store, and the old HD skins do not fit.

2022-06-06 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2022-06-03 · open the original ↗

• The clip contains no substantive content. The captions for this 10-second short consist only of the words "wait a minute".

2022-05-31 · open the original ↗

• The host calls the Pro a worthwhile upgrade, mainly for package detection. He thinks package alerts justify the extra cost, especially where porch pirates are common, and also points to USB-C power and future integrations. A doorbell press reached his phone in about one second, and a package alert in under 30 seconds.
• Package alerts come only from the downward package camera. That camera is 8 MP and records 1600x1200 at 2 frames per second. In the host's test, a package the main camera could see but the package camera could not raised no alert, and one only the package camera could see did. He calls this an oversight.
• Package detection in UniFi Protect 2.0 looks half-implemented. The non-Pro G4 Doorbell offers package smart detection but has no package notification option. Protect 2.0 also will not let one device sit in more than one device group. The animated display shown in the launch video was missing as of Protect 2.0.
• Image quality matches the original G4 Doorbell, below Ubiquiti's standalone cameras. The host believes both doorbells use the same 5 MP main camera. The view is very wide and displayed at 1600x1200, and shots are backlit on default settings, but a year with the original never left him unable to tell who was at the door.
• The speaker is still too quiet for two-way talk. Microphone gain seems better than on the G4 Doorbell, but the host says he will never use the talk feature, since couriers have usually gone by the time the app opens. The fingerprint and NFC reader did nothing yet, and the rated operating ceiling of 40 °C (104 °F) worries him for doors in direct sun.

2022-05-30 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2022-05-26 · open the original ↗

• Interviewers pick questions by assigned competency and the candidate's resume. Zhang, a former Amazon software engineer who says he conducted a couple dozen coding interviews, describes a 10–15 minute introduction, one or two technical questions from a pool (a second only if the first goes well), then a few minutes of Q&A. Design interviews are different: talking- and drawing-heavy, judged on what the candidate says and asks while producing a scalable, highly available distributed design.
• Follow-up questions expose candidates who only memorized practice problems. After the code is written, usually in a plain editor and never run, he asks why a variable or approach was chosen; drilled candidates "typically don't know what to answer". A struggling candidate gets hints, and he expects a good developer to reach the answer with them; questions are sometimes left deliberately vague so the candidate will ask clarifying questions.
• Cultural-fit answers need concrete, real examples. He asks how candidates handle conflict or a missed deadline and checks the answer against company values; vague examples, when probed, reveal themselves as made up. Giard offers his own conflict answer, that two people who disagree can both be right and not every dispute is "a hill I want to die on".
• Never ask the interviewer for the answer to the last question. Zhang says it makes things awkward and that interviewers are legally not allowed to tell, since revealing it and then rejecting the candidate could cause trouble. Better questions for the closing minutes are about the interviewer's day-to-day and what they like or dislike about the job.
• Preparation is a muscle built over time, not in a week. He recommends practising the widely published data-structure and algorithm questions on leetcode.com and doing mock interviews with friends to practise explaining code, since perfect code you cannot explain does not make a good interview. He distrusts online offers to prepare you in a week or a day.

2022-05-26 · open the original ↗

• Word's checker can run on spelling alone or spelling and grammar. The choice is a dropdown on the Spelling and Grammar button on the Review ribbon; Giard always leaves it on both, and notes that F7, though not shown on the button, has long been the shortcut. The dialog that opens on the right suggests a correction or lets you Ignore or Ignore All (for example a proper name you don't want in your dictionary), and in his demo it flagged a verb that did not match its subject ("shows" where "show" was needed).
• Proofing language sets which variant's spelling rules apply. Under Review, the Language button's dropdown has Set Proofing Language, which Giard says he has traditionally used to set English (United States). It lists regional variants, such as several dialects of Arabic and UK or Australian English, which decide whether a word like "flavor" ends in "-or" or "-our".
• Marked sections can be excluded from spelling and grammar checks. Highlight the text, open Set Proofing Language, and tick "Do not check spelling or grammar"; Giard says he only recently learned of this option. He applies it to the code samples in a technical article (variable, table and function names that are not English words), and the check then skips them, not stopping on identifiers like "P1" or "pack_array". The setting is saved with the document and applies the next time it is opened and checked.

2022-05-25 · open the original ↗

• Logs alone cannot trace one request across a distributed system. Wijesiriwardena, six weeks into Microsoft's CSE team after Microsoft consultancy, says not every system feeds one log aggregator, and filtering out a single request needs a correlation ID, neither of which is trivial in production. Distributed tracing isolates one request from possibly millions, with its context and dependencies.
• OpenTelemetry is a set of standards, SDKs and tools. The standard defines a trace ID with bytes reserved for parent and child, which keeps dependency relationships as a request moves through the system, and builds on the W3C trace context headers for context propagation. HTTP is not required: popular libraries are already instrumented, but over a custom protocol or a message queue you insert and extract the trace context and "baggage" yourself.
• Its job ends at delivering telemetry to your ops team's tool. A collector runs in-process or as a sidecar and forwards data, through an exporter, to an endpoint such as Azure Monitor, Jaeger or Prometheus; OpenTelemetry says nothing about querying or reporting. He argues the richer data it supplies is what is driving new query and visualization features in those tools.
• Third-party libraries can opt in rather than stay black boxes. He says a SQL Server driver or the .NET HTTP client can be told to emit its internal telemetry the OpenTelemetry way. To learn, he points to the OpenTelemetry GitHub's per-language SDK samples, including custom instrumentation for RabbitMQ, and the CNCF Slack's OpenTelemetry channel; documentation is where he sees most scope for contribution.
• Observability matters in 2022 because of a second cloud-migration wave. He reports what he hears in the industry: ten years ago applications were moved whole onto VMs, and now they are being broken into distributed systems owned by many teams on mixed technology stacks, so a single view of telemetry is needed to diagnose problems quickly.

2022-05-23 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2022-05-21 · open the original ↗

• A 30-year-old Access 2.0 database still accepts new data. One speaker shows Microsoft Access 2.0, which Julia worked on, on a period machine with 16 MB, a 120 MB hard drive and a 3.5-inch floppy; to get a 21K, 16-colour picture of her into it they went through a floppy-drive-to-USB-A-to-USB-C adapter chain, dithered the image to 16 colours and embedded it in the database alongside the other "employees of the day".
• The tribute misspelled her name, which landed as the punchline. The other voice points out the last name is wrong, and the first jokes about fixing it in the database.

2022-05-21 · open the original ↗

• The archive vault prizes constant conditions over cold. The archivist (addressed as Amy) says the vault is held at 55 degrees and 45% humidity for storing CD-ROM media, and that the constancy matters more than the temperature itself.
• The collection spans shipped media, hardware and branded oddities. Hanselman is shown 8-, 5- and 3.5-inch floppies, a boxed Excel, Flight Simulator for Windows 95 (the first Windows version; earlier ones were DOS), a SideWinder controller, an internet keyboard and the Surface table. The archivist calls Clippy the most popular item, lent out as a costume, and describes animatronic DW and Arthur toys that talked and sang along with the TV show via a transponder, with add-on packs for VHS or PC.
• Paper tape beat punch cards because a drop didn't destroy the program. The archivist shows a hex assembler shipped on paper tape, the way software was distributed before disks.
• Original Xbox dev kits survive because an Office admin bought them cheaply. Before the console design was shown, the dev kits came in PC-style boxes; an admin found she could order them internally for $45 each, and her team wiped the Xbox OS, installed Windows and used them as desktops.
• The archive collects partly by automatic copies, partly by donation. Internal posters are sent to it automatically, while external marketing material only arrives if teams hand it over; an Apple II on a cart is part of a larger donation meant to go on to Stanford. A rack system that would have cost $1–8 million in 2003 was, by 2008, worth so little that HP paid two people two days to dismantle and rebuild it there. The archivist says a volunteer reconditioning programme for old machines is likely, and that outside researchers get in only if sponsored, as it is not open to the public.

2022-05-19 · open the original ↗

• A basic Omada home network needs three separated LANs and SSIDs. The host builds it from factory defaults on a TL-R605 router, a TL-SG2210MP PoE switch, an in-wall Omada access point and an OC200 controller, which a software controller can replace. The secure LAN is 192.168.42.1/24, guest is VLAN 10 (192.168.10.x) and IoT is VLAN 107 (192.168.107.x).
• Moving off the default subnet strands devices on old DHCP leases. The host avoids 192.168.0.x out of habit, not for security, starts DHCP at .10 to keep .2–.9 free for static addresses, and sets DNS to 1.1.1.1 and 9.9.9.9, or a Pi-hole's address. Rather than wait for leases to expire, he power-cycled the PoE switch. The controller then needed a reboot before it took its fixed IP of .2.
• Guest Wi-Fi uses Omada's guest network policy plus a rate limit. The policy blocks guest clients from every private IP subnet but still lets them reach the internet, and the SSID is tagged VLAN 10. A rate-limit profile caps guests at 20 Mbps down and 5 Mbps up so they cannot use up the household's bandwidth.
• The IoT network relies on a one-way switch ACL, not guest isolation. Guest isolation would block casting between IoT devices, so the host instead denies all protocols from the IoT network to the secure network, leaving the rule one-way. Ping tests confirmed IoT was blocked from secure but kept internet access. He puts his own phone on the IoT network to cast rather than writing cross-network ACLs.

2022-05-17 · open the original ↗

• Newer UniFi PoE cannot power the NanoBeam 5AC directly. The $99 NanoBeam runs on 24 V passive PoE, while the host's UniFi Lite 8 PoE switch and UDM SE supply only 802.3af. He bridges the gap with an INS-3AF-I-G adapter; the bundled injector or the older US-8-150W, which is selectable per port, also work. The side where the network originates is the access point, and the far ends are stations.
• Each unit starts at 192.168.1.20 and must be re-addressed. The host gives his laptop a static 192.168.1.99 and moves the units to 192.168.42.10, .11 and .12 so they can reach the internet and update from airOS 8.7.1 to 8.7.9. He disables the Bluetooth management radio and saves a configuration backup of every unit, which the setup wizard can restore onto a replacement.
• Narrow the radio to what the site needs. The host uses 80 MHz channels instead of the default 40 MHz only because his rural site is quiet; wider channels give more bandwidth but pick up more interference. He removes DFS channels, which give way to priority users such as airports and the military, and sets 0 dBm output on the bench out of a −4 to 21 range.
• Update firmware on stations first, then the access point. If the access point is updated first and the versions turn out incompatible, remote stations become unreachable without a site visit. In this test an 8.7.1 station linked to an 8.7.9 access point, which the host says used not to work. Lying flat on his desk, the link reported 250–300 Mbps of capacity and a 77/41 Mbps speed test, and he expects about 300 Mbps once aimed.
• Point-to-multipoint requires turning off airOS 8's point-to-point mode. With that mode on at the access point, a second station could see the SSID but not select it. After the host turned it off, station one did not reconnect within five minutes until he also unchecked the mode on the station and reselected the SSID. Point-to-point mode accepts only airMAX AC stations, while mixed mode also takes older M5 units.

2022-05-16 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.