知玄｜盈尺 surinch

PostgreSQL 14 hits EOL on Nov 12. We tested 14→17: pg_upgrade took ~2s, but pg_stats was empty afterwards and one JOIN went 184ms → 994ms. Run vacuumdb --analyze-in-stages BEFORE sending traffic. Upgrading to 18? Re-create extended stats. PG14 将于 11-12 停更。实测 14→17：pg_upgrade 只用 2 秒，但升完统计信息为空，同一条 JOIN 从 184ms 变成 994ms。放流量前先跑 vacuumdb --analyze-in-stages；升 18 记得补扩展统计。 PG14 升级体检 11-11 上线，前 50 位预约送迁移风险评估： surinch.com/tools/db-migrati… #PostgreSQL

蓝v流量也没变化嘛

Amazon FBA sellers heading into peak: what's your rule for aged inventory this month? A) remove/dispose before Nov 14 (skip this month's aged surcharge) B) keep selling through BFCM and eat the surcharge + peak storage C) liquidate / Grade and Resell first, remove the rest Which one, and why? 旺季库龄怎么处理：11-14 前移除、硬扛附加费卖到黑五，还是先清货再移除？你怎么选？ #AmazonFBA

9810 overseas-warehouse pre-refund (China tax): the paperwork is a mapping problem, not a rate problem. Keep four IDs tied together for every batch: customs item number ↔ your SKU ↔ FBA shipment ID ↔ sales record. Without that chain, the "overseas warehouse use" proof and the later reconciliation fall apart. We don't publish rates or refund percentages here — ask your tax authority. The process and the evidence list are what you can control. 9810 海外仓预退税：先把项号—SKU—FBA 货件—销售记录四条线对齐，再谈单证。税率和比例以主管税务机关为准。 #AmazonSeller

SP-API Fulfillment Inbound (v2024-03-20): generatePlacementOptions and setTransport are redoable. confirmPlacementOption is not. Amazon's own docs: once you confirm a placement option, it cannot be reversed. Partner-carrier quotes expire (small parcel ~24h, LTL/FTL ~1h). Changing box contents means regenerating placement options. Automate the comparison table before confirm. Leave the confirm click to a human. SP-API 建 FBA 货件：比对表可以自动化，confirmPlacementOption 不要交给脚本。确认分仓不可撤销。 #AmazonFBA #SPAPI

Amazon Seller Assistant (and the Claude selling-partner plugin) can change listings, ads and account settings in a few clicks. Before you let it, write the rules down: 1. Read-only by default. Writes need a human yes. 2. Quote the Seller Central line or report row it used. No quote → reject. 3. Never touch irreversible steps: confirmPlacementOption, carrier acceptance with a fee clock, account appeals that admit fault. 4. One change per run. Batch edits are how silent damage spreads. We keep the same four rules for our own AI ops scripts. Template and checklist: surinch.com/resources/amazon… #AmazonSeller #AI

Got Amazon's "enhanced identity video verification" email? Treat it as a document request, not a Plan of Action. Official replies on the forums: book within 14 days, bring the same ID and bank statement you used to register (no new docs), about 15 minutes, camera + mic required. Don't wait until day 13, and don't swap the legal person just for the call — they won't accept new identity papers. If you haven't been asked, ignore third-party "wave" posts. Act only on the email in your Account Health queue. 增强型视频验证：当资料请求，别写 POA。14 天内预约，只认注册时证件和银行对账单原件，约 15 分钟。别拖到第 13 天，也别临时换法人。步骤与注意点： surinch.com/resources/amazon… #AmazonSeller

FBA tip: submit removal/disposal orders by Nov 14, 11:59pm PT and those units skip this month's aged-inventory surcharge (assessed on the 15th), even if they haven't left the warehouse yet. Peak storage (~$2.40/cu ft) is billed on top. Sort SKUs by volume × age first. FBA 超龄附加费每月 15 日评估。11-14 23:59（太平洋时间）前提交移除/弃置，这些件数当月就不再计费，哪怕货还没出仓。旺季仓储约 $2.40/立方英尺另算。先按体积×库龄排序，再决定卖、移还是弃： surinch.com/resources/amazon… #AmazonFBA

PostgreSQL 14 reaches end of life on Nov 12, one week from today. Still running it somewhere? What's your path: pg_upgrade --link, logical replication to a new cluster, or a dump/restore window? And are you going to 17 or straight to 18? PG14 下周四（11-12）停更。你还有 14 在跑吗？准备怎么升：pg_upgrade、逻辑复制还是停机导入？升 17 还是直接 18？ #PostgreSQL

Postgres table bloating even though autovacuum runs all day? Look for an old transaction. VACUUM can only remove dead rows that no open transaction can still see. A session sitting idle in transaction after a write, or holding a snapshot (REPEATABLE READ, an open cursor), pins that horizon, and cleanup stalls across the database. Find the oldest ones: SELECT pid, usename, state, now() - xact_start AS xact_age, backend_xid, backend_xmin FROM pg_stat_activity WHERE backend_xid IS NOT NULL OR backend_xmin IS NOT NULL ORDER BY xact_start; Then add a guardrail: ALTER ROLE app SET idle_in_transaction_session_timeout = '5min'; Also check pg_replication_slots and pg_prepared_xacts. Stale slots and forgotten prepared transactions block cleanup the same way. #PostgreSQL

MySQL ERROR 1146 "Table 'x.y' doesn't exist", but you can see the table? Two usual suspects: 1. Wrong database. The name in the error is the schema MySQL actually searched: your connection's default DB if you didn't qualify the table. 2. Case. On Linux, lower_case_table_names defaults to 0, so Orders and orders are different tables. Windows defaults to 1 and macOS to 2, which is why SQL that worked on a laptop fails on the server. It can only be set when the data directory is initialized. One query checks both: SELECT TABLE_SCHEMA, TABLE_NAME FROM information_schema.TABLES WHERE LOWER(TABLE_NAME) = LOWER('Orders'); Bonus trap: `db.Orders` inside one pair of backticks is a single table name. Write `db`.`Orders`. #MySQL

Upgraded MySQL 8.0 → 8.4 and bulk writes got 2.3–3.4x slower? Check innodb_change_buffering first. 8.4 changed its default from all to none. In our lab (8.0.46 vs 8.4.10, synthetic data, secondary indexes bigger than a 128MB buffer pool), writing 200k rows: • 8.0 defaults: 4.91–5.20 s • 8.4 defaults: 12.09–16.55 s • 8.4 + innodb_change_buffering=all: 6.30–6.46 s • 8.4 defaults + 1GB buffer pool: 3.14–3.32 s Tell-tale sign: Innodb_buffer_pool_reads jumps (~206k per run vs ~5k with change buffering back on). The adaptive hash index (also OFF by default in 8.4) wasn't the cause. It only mattered for 8-way concurrent joins (~28%). Don't paste your whole 8.0 my.cnf back. Pick per workload: change buffering on, or enough memory for the indexes. MySQL 8.4 升级后写入变慢，主因是 change buffering 默认改成 none。完整数据、13 项默认值对照和只读排查 SQL： surinch.com/resources/mysql-… #MySQL

MySQL ERROR 1452: Cannot add or update a child row: a foreign key constraint fails. The value you're writing into the child column has no matching row in the parent table. The full message names the constraint, the child column and the parent column, so read it before guessing. Find the orphans first (read-only): SELECT c.parent_id FROM child c LEFT JOIN parent p ON p.id = c.parent_id WHERE c.parent_id IS NOT NULL AND p.id IS NULL; Careful with SET FOREIGN_KEY_CHECKS=0 during imports: turning it back on does not re-check rows already loaded, so orphans stay in the table silently. #MySQL

Amazon sellers using ChatGPT, Claude or DeepSeek day to day: what's one task you now let it do without checking the output? And one you stopped trusting it with? 用 AI 做亚马逊运营的朋友：哪件事你已经放心交给它、不再复核？哪件事试过之后不敢再交给它？ #AmazonSeller #AI

Using AI to answer Amazon Account Health notices? Make it classify before it writes. Most notices need one of four replies: fix the listing, send documents, write a Plan of Action, or nothing at all. A model's default is a polite, apologetic POA for everything. That's the wrong document when the fix is a listing edit, or when the complaint has already been withdrawn. Our order: AI reads the notice, picks the reply type and quotes the line that justifies it. A human checks that line. Only then does AI draft. AI 回亚马逊绩效通知：先让它判断该怎么回（改 Listing、交文件、写 POA、不用回），并引用通知原文作依据，人核对后再写。什么都写道歉 POA 是最常见的错。 #AmazonSeller #AI

Shipping FBM from outside the EU and Buy Shipping says "Required attribute missing ITEM_MANUFACTURER_PRODUCT_ID"? That's the EU product-ID rule, mandatory since Nov 1 for B2C parcels up to €150. Each item needs a real model or part number. What doesn't count, per Amazon's SP-API answer: • your seller SKU • "N/A" • the value in your listing: Buy Shipping doesn't pull it, you still enter it per item GTIN-exempt brands get hit hardest. In our own 22-SKU snapshot, only 7 had a model or part number that wasn't just the SKU. 欧盟 11-01 起强制商品编码。FBM 用 Buy Shipping 报 ITEM_MANUFACTURER_PRODUCT_ID，就是缺真实型号或零件号，填 SKU、N/A 都不算。三种编码对照和修法： surinch.com/resources/amazon… #AmazonSeller

Amazon listing error 8541, decoded: the UPC/EAN you submitted already matches an ASIN, and some of your values (title, brand, size, color…) conflict with the catalog. Amazon keeps the catalog data and rejects yours. Work it in this order: 1. Open the ASIN the error names. Is it really the same product? 2. Same product: match your file to the catalog values, or drop the conflicting attributes and list against that ASIN 3. Different product: the barcode is wrong for it. Check GS1 before anything else 4. You own the brand and the catalog is wrong: fix it through Brand Registry, not a new listing 8541：你的条码已对应某个 ASIN，但标题、品牌、尺寸等和目录冲突，亚马逊以目录为准。先确认是不是同一个商品，再决定对齐目录还是查条码。 #AmazonSeller

Amazon US auto-created Virtual Multipacks (SKUs starting with VMP_) for some brand-owned FBA items. Review, reprice or delete by Nov 10; the rest go live Nov 11. Default price = unit price × qty. Search VMP_ in Manage All Inventory today. 亚马逊美国站已给部分品牌 FBA 商品自动建好虚拟多件装（SKU 以 VMP_ 开头）。11-10 前可审核、改价或删除，11-11 起没删的直接开售。默认价=单件价×件数，履约费据称也按件数乘，薄利款要先算账。今天就去 Manage All Inventory 搜 VMP_： surinch.com/resources/amazon… #AmazonFBA

Amazon EU FBA sellers: since Nov 2 the fuel & logistics surcharge is 2% of the fulfilment fee (was 1.5%) in UK, DE, FR, IT, ES, PL, SE, NL, IE and BE. MCF follows on Nov 19 in UK, DE, FR, IT and ES. It's charged on the fee, not your price. DE small parcel ≤400g at the €3.60 peak fee: +€0.018 per unit. On a €9.99 item that's about 0.2 margin points. Our take: don't reprice for this alone. Use it as the reason to re-pull Fee Preview, check repricer floors that hard-code old fees, and quote MCF orders correctly after Nov 19. 欧洲站 FBA 燃油与物流附加费 11-02 起 1.5%→2%，按配送费算、不按售价，每件约多 €0.02。不用单独调价，顺手查一下自动调价的最低价是不是写死了旧费用。 #AmazonFBA

Postgres tip: EXPLAIN ANALYZE shows "Sort Method: external merge Disk: …"? The sort didn't fit in work_mem and spilled to disk. Don't fix it by raising work_mem globally. It's a limit per sort or hash operation, not per connection, so one complex query can use it several times over, on every connection. Hash nodes can use work_mem × hash_mem_multiplier (default 2.0 since PG15). Raise it only where it's needed: BEGIN; SET LOCAL work_mem = '256MB'; -- the report query COMMIT; Then re-run EXPLAIN ANALYZE and look for "Sort Method: quicksort Memory: …". #PostgreSQL