The direct answer
No. Keep the first slice read-only or in shadow until you have compared committed outcomes from the old and new PostgreSQL paths under a concurrent workload, observed outstanding locks, and tested a rollback. Single-threaded parity cannot settle this, because the anomalies that break multi-step invariants only appear when transactions overlap. This is a proposed gate, not a claim that any particular slice is safe; the PostgreSQL documentation describes the database's behavior, not your schema or workload.
Why single-threaded parity is weak evidence
Fowler's Strangler Fig essay, published on his site in a 22 August 2024 revision of the 2004 piece, describes gradual modernization as building small additions on top of, yet separate from, the legacy code base, then moving bits of behavior across. He warns that wholesale replacement plans that sound simple have gone down in flames most of the time, because replacement takes a long time, users cannot wait for new features, and existing behavior is hard to specify, including behavior nobody wants. Treat the first slice as an evidence-generating step, and define the gate before the slice carries traffic; that framing is our proposal built on Fowler's, not his conclusion about your project.
The PostgreSQL Transaction Isolation documentation supplies the mechanism. Read Committed is the default isolation level, and the documentation's Table 13.1 shows it permits nonrepeatable reads, phantom reads and serialization anomalies. The documentation also states that two successive SELECT commands inside one transaction can see different data if other transactions commit between them. Our inference from that documented behavior: a parity test that runs one operation at a time does not exercise these phenomena, so passing it is weak evidence about whether the new path preserves an invariant that spans multiple rows or tables. This is a reasoned inference from the isolation table, not a measured failure rate.
The isolation and lock evidence to collect
The same documentation table narrows the choice: PostgreSQL's Repeatable Read does not allow phantom reads but still permits serialization anomalies, and Serializable permits none of the listed phenomena. If the slice must preserve an invariant across concurrent transactions, the team should test whether Serializable isolation or explicit locking is required. That is a proposal; the PostgreSQL documents describe the mechanisms but do not determine the right choice for your schema.
For locking, the PostgreSQL Explicit Locking documentation states that two transactions cannot hold locks of conflicting modes on the same table at the same time, and that UPDATE, DELETE, INSERT and MERGE acquire ROW EXCLUSIVE locks on the target table, while SELECT acquires ACCESS SHARE, which conflicts only with ACCESS EXCLUSIVE. The documentation names pg_locks as the system view for examining the list of currently outstanding locks. A proposed pre-cutover check is to observe outstanding locks and their granted status under the concurrent workload and agree a threshold before the cutover; whether pg_locks also exposes wait timing in your PostgreSQL version is something to verify in the full documentation rather than assume here.
Name the exact migration subform
The PostgreSQL ALTER TABLE documentation lists separate subforms in its synopsis: adding a column, dropping a column, changing a column type, and adding or altering constraints are distinct operations in the command's grammar. Our inference from that synopsis: a cutover migration plan should name the exact subform it will run, because treating ALTER TABLE as one undifferentiated operation hides differences the documentation itself distinguishes. Verify the lock behavior of your specific subform against the full PostgreSQL documentation before the change window; this article does not assign lock levels to subforms.
A proposed pre-cutover evidence packet
The following is a proposed diligence method, built from Fowler's incremental-modernization frame and PostgreSQL's documented behavior, not a claim about your system. First, an operation inventory: list every operation the slice will handle, and for each one that writes more than one row, record the transaction isolation level and any explicit locking. Second, a concurrent workload: run the old and new paths against the same overlapping transactions, not one at a time. Third, a committed-outcome comparison: compare final committed state, not only individual responses, and explain every difference. Fourth, lock observation through pg_locks against the agreed threshold. Fifth, a rollback condition: a tested path back to the legacy route, written down before cutover, with the event that triggers it.
Fowler lists four high-level activities for this kind of incremental approach: understand the outcomes you want, decide how to break the problem into smaller parts, successfully deliver the parts, and change the organization to allow this to happen on an ongoing basis. He notes the ordering does not imply a sequence. The evidence packet is the concrete form of the third activity for a first slice. Fowler's excerpt ties lower risk to the components being small: because the replaced pieces are small, he writes, there is not so much risk involved when the new software is introduced. The rollback condition is our proposal for making the slice recoverable, not a point Fowler makes; keep the two ideas separate in the plan.
Who owns each step
The gate has three owners regardless of staffing model. Someone owns the operation inventory and the isolation decision; someone owns running the concurrent workload and explaining differences; someone owns the rollback decision during the cutover window. Name all three before the slice starts.
If you staff this with a managed team run by our tech lead, our tech lead talks with you every day, turns your priorities into a plan, owns code review, quality and releases, and manages the engineers; you own the priorities and Atom owns delivery. It is priced per team per month: Starter at 29,500 USD, Core at 44,500 USD, Plus at 67,500 USD, with an additional engineer at 10,900 USD senior or 8,000 USD mid. It ends with 30 days' notice to remove one person and 60 days to end the team, and if the first sprint misses its agreed goal and you end the team, sprint one is free.
If you staff it with developers in your team, our engineers join your standups, tools and repositories and take direction from your engineering manager, so you direct the work and own delivery quality. That is priced per developer per month, full time at about 160 hours: 7,000 USD mid, 9,500 USD senior, 11,500 USD senior data/AI, 12,500 USD lead, 6,500 USD QA automation, with a 3-month initial term and then 30 days' notice. If nobody on your side will review code, plan the cutover window and own the rollback call, buy the tech lead, not the seats. If you have no written scope yet, the Blueprint is a two-week, fixed-price 7,500 USD engagement that produces the scope, architecture, milestone plan, team plan and a fixed monthly quote; you keep it and it is not credited toward a team.
Conditions that change the recommendation
The recommendation to hold at read-only or shadow weakens under specific conditions, and you should check them rather than assume. Our inference: if the slice writes exactly one row per transaction with no cross-row invariant, the concurrency exposure the isolation table describes is smaller, though the lock observation still applies. If your workload is genuinely single-writer, the concurrent-parity step may be shortened, but verify that claim against your actual traffic rather than an assumption.
Conversely, if the slice touches tables that concurrent reporting or batch jobs also use, be precise about what conflicts. The PostgreSQL Explicit Locking documentation lists ROW EXCLUSIVE as conflicting with SHARE, SHARE ROW EXCLUSIVE, EXCLUSIVE and ACCESS EXCLUSIVE, not with another ROW EXCLUSIVE, so two ordinary writers do not conflict at table level; their contention is at row level. Table-level conflict comes from operations that take the stronger modes the documentation names, such as CREATE INDEX without CONCURRENTLY, which acquires SHARE, or TRUNCATE, which the documentation says obtains ACCESS EXCLUSIVE. Check which of those run against the slice's tables during the shadow period, and lengthen it if they do. And if the committed-outcome comparison shows differences you cannot explain, that is not a tuning problem; it is evidence the slice does not yet replicate legacy behavior, and the cutover should not proceed until it is explained.
